fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/statfs.h>
  34 #include <linux/compat.h>
  35 #include <linux/bit_spinlock.h>
  36 #include <linux/xattr.h>
  37 #include <linux/posix_acl.h>
  38 #include <linux/falloc.h>
  39 #include <linux/slab.h>
  40 #include <linux/ratelimit.h>
  41 #include <linux/mount.h>
  42 #include <linux/btrfs.h>
  43 #include <linux/blkdev.h>
  44 #include <linux/posix_acl_xattr.h>
  45 #include <linux/uio.h>
  46 #include "ctree.h"
  47 #include "disk-io.h"
  48 #include "transaction.h"
  49 #include "btrfs_inode.h"
  50 #include "print-tree.h"
  51 #include "ordered-data.h"
  52 #include "xattr.h"
  53 #include "tree-log.h"
  54 #include "volumes.h"
  55 #include "compression.h"
  56 #include "locking.h"
  57 #include "free-space-cache.h"
  58 #include "inode-map.h"
  59 #include "backref.h"
  60 #include "hash.h"
  61 #include "props.h"
  62 #include "qgroup.h"
  63
  64 struct btrfs_iget_args {
  65         struct btrfs_key *location;
  66         struct btrfs_root *root;
  67 };
  68
  69 struct btrfs_dio_data {
  70         u64 outstanding_extents;
  71         u64 reserve;
  72         u64 unsubmitted_oe_range_start;
  73         u64 unsubmitted_oe_range_end;
  74 };
  75
  76 static const struct inode_operations btrfs_dir_inode_operations;
  77 static const struct inode_operations btrfs_symlink_inode_operations;
  78 static const struct inode_operations btrfs_dir_ro_inode_operations;
  79 static const struct inode_operations btrfs_special_inode_operations;
  80 static const struct inode_operations btrfs_file_inode_operations;
  81 static const struct address_space_operations btrfs_aops;
  82 static const struct address_space_operations btrfs_symlink_aops;
  83 static const struct file_operations btrfs_dir_file_operations;
  84 static const struct extent_io_ops btrfs_extent_io_ops;
  85
  86 static struct kmem_cache *btrfs_inode_cachep;
  87 struct kmem_cache *btrfs_trans_handle_cachep;
  88 struct kmem_cache *btrfs_transaction_cachep;
  89 struct kmem_cache *btrfs_path_cachep;
  90 struct kmem_cache *btrfs_free_space_cachep;
  91
  92 #define S_SHIFT 12
  93 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  94         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  95         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  96         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  97         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  98         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  99         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 100         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 101 };
 102
 103 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 104 static int btrfs_truncate(struct inode *inode);
 105 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 106 static noinline int cow_file_range(struct inode *inode,
 107                                    struct page *locked_page,
 108                                    u64 start, u64 end, int *page_started,
 109                                    unsigned long *nr_written, int unlock);
 110 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
 111                                            u64 len, u64 orig_start,
 112                                            u64 block_start, u64 block_len,
 113                                            u64 orig_block_len, u64 ram_bytes,
 114                                            int type);
 115
 116 static int btrfs_dirty_inode(struct inode *inode);
 117
 118 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 119 void btrfs_test_inode_set_ops(struct inode *inode)
 120 {
 121         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 122 }
 123 #endif
 124
 125 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 126                                      struct inode *inode,  struct inode *dir,
 127                                      const struct qstr *qstr)
 128 {
 129         int err;
 130
 131         err = btrfs_init_acl(trans, inode, dir);
 132         if (!err)
 133                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 134         return err;
 135 }
 136
 137 /*
 138  * this does all the hard work for inserting an inline extent into
 139  * the btree.  The caller should have done a btrfs_drop_extents so that
 140  * no overlapping inline items exist in the btree
 141  */
 142 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 143                                 struct btrfs_path *path, int extent_inserted,
 144                                 struct btrfs_root *root, struct inode *inode,
 145                                 u64 start, size_t size, size_t compressed_size,
 146                                 int compress_type,
 147                                 struct page **compressed_pages)
 148 {
 149         struct extent_buffer *leaf;
 150         struct page *page = NULL;
 151         char *kaddr;
 152         unsigned long ptr;
 153         struct btrfs_file_extent_item *ei;
 154         int err = 0;
 155         int ret;
 156         size_t cur_size = size;
 157         unsigned long offset;
 158
 159         if (compressed_size && compressed_pages)
 160                 cur_size = compressed_size;
 161
 162         inode_add_bytes(inode, size);
 163
 164         if (!extent_inserted) {
 165                 struct btrfs_key key;
 166                 size_t datasize;
 167
 168                 key.objectid = btrfs_ino(inode);
 169                 key.offset = start;
 170                 key.type = BTRFS_EXTENT_DATA_KEY;
 171
 172                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 173                 path->leave_spinning = 1;
 174                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 175                                               datasize);
 176                 if (ret) {
 177                         err = ret;
 178                         goto fail;
 179                 }
 180         }
 181         leaf = path->nodes[0];
 182         ei = btrfs_item_ptr(leaf, path->slots[0],
 183                             struct btrfs_file_extent_item);
 184         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 185         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 186         btrfs_set_file_extent_encryption(leaf, ei, 0);
 187         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 188         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 189         ptr = btrfs_file_extent_inline_start(ei);
 190
 191         if (compress_type != BTRFS_COMPRESS_NONE) {
 192                 struct page *cpage;
 193                 int i = 0;
 194                 while (compressed_size > 0) {
 195                         cpage = compressed_pages[i];
 196                         cur_size = min_t(unsigned long, compressed_size,
 197                                        PAGE_SIZE);
 198
 199                         kaddr = kmap_atomic(cpage);
 200                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 201                         kunmap_atomic(kaddr);
 202
 203                         i++;
 204                         ptr += cur_size;
 205                         compressed_size -= cur_size;
 206                 }
 207                 btrfs_set_file_extent_compression(leaf, ei,
 208                                                   compress_type);
 209         } else {
 210                 page = find_get_page(inode->i_mapping,
 211                                      start >> PAGE_SHIFT);
 212                 btrfs_set_file_extent_compression(leaf, ei, 0);
 213                 kaddr = kmap_atomic(page);
 214                 offset = start & (PAGE_SIZE - 1);
 215                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 216                 kunmap_atomic(kaddr);
 217                 put_page(page);
 218         }
 219         btrfs_mark_buffer_dirty(leaf);
 220         btrfs_release_path(path);
 221
 222         /*
 223          * we're an inline extent, so nobody can
 224          * extend the file past i_size without locking
 225          * a page we already have locked.
 226          *
 227          * We must do any isize and inode updates
 228          * before we unlock the pages.  Otherwise we
 229          * could end up racing with unlink.
 230          */
 231         BTRFS_I(inode)->disk_i_size = inode->i_size;
 232         ret = btrfs_update_inode(trans, root, inode);
 233
 234         return ret;
 235 fail:
 236         return err;
 237 }
 238
 239
 240 /*
 241  * conditionally insert an inline extent into the file.  This
 242  * does the checks required to make sure the data is small enough
 243  * to fit as an inline extent.
 244  */
 245 static noinline int cow_file_range_inline(struct btrfs_root *root,
 246                                           struct inode *inode, u64 start,
 247                                           u64 end, size_t compressed_size,
 248                                           int compress_type,
 249                                           struct page **compressed_pages)
 250 {
 251         struct btrfs_trans_handle *trans;
 252         u64 isize = i_size_read(inode);
 253         u64 actual_end = min(end + 1, isize);
 254         u64 inline_len = actual_end - start;
 255         u64 aligned_end = ALIGN(end, root->sectorsize);
 256         u64 data_len = inline_len;
 257         int ret;
 258         struct btrfs_path *path;
 259         int extent_inserted = 0;
 260         u32 extent_item_size;
 261
 262         if (compressed_size)
 263                 data_len = compressed_size;
 264
 265         if (start > 0 ||
 266             actual_end > root->sectorsize ||
 267             data_len > BTRFS_MAX_INLINE_DATA_SIZE(root) ||
 268             (!compressed_size &&
 269             (actual_end & (root->sectorsize - 1)) == 0) ||
 270             end + 1 < isize ||
 271             data_len > root->fs_info->max_inline) {
 272                 return 1;
 273         }
 274
 275         path = btrfs_alloc_path();
 276         if (!path)
 277                 return -ENOMEM;
 278
 279         trans = btrfs_join_transaction(root);
 280         if (IS_ERR(trans)) {
 281                 btrfs_free_path(path);
 282                 return PTR_ERR(trans);
 283         }
 284         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 285
 286         if (compressed_size && compressed_pages)
 287                 extent_item_size = btrfs_file_extent_calc_inline_size(
 288                    compressed_size);
 289         else
 290                 extent_item_size = btrfs_file_extent_calc_inline_size(
 291                     inline_len);
 292
 293         ret = __btrfs_drop_extents(trans, root, inode, path,
 294                                    start, aligned_end, NULL,
 295                                    1, 1, extent_item_size, &extent_inserted);
 296         if (ret) {
 297                 btrfs_abort_transaction(trans, root, ret);
 298                 goto out;
 299         }
 300
 301         if (isize > actual_end)
 302                 inline_len = min_t(u64, isize, actual_end);
 303         ret = insert_inline_extent(trans, path, extent_inserted,
 304                                    root, inode, start,
 305                                    inline_len, compressed_size,
 306                                    compress_type, compressed_pages);
 307         if (ret && ret != -ENOSPC) {
 308                 btrfs_abort_transaction(trans, root, ret);
 309                 goto out;
 310         } else if (ret == -ENOSPC) {
 311                 ret = 1;
 312                 goto out;
 313         }
 314
 315         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 316         btrfs_delalloc_release_metadata(inode, end + 1 - start);
 317         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
 318 out:
 319         /*
 320          * Don't forget to free the reserved space, as for inlined extent
 321          * it won't count as data extent, free them directly here.
 322          * And at reserve time, it's always aligned to page size, so
 323          * just free one page here.
 324          */
 325         btrfs_qgroup_free_data(inode, 0, PAGE_SIZE);
 326         btrfs_free_path(path);
 327         btrfs_end_transaction(trans, root);
 328         return ret;
 329 }
 330
 331 struct async_extent {
 332         u64 start;
 333         u64 ram_size;
 334         u64 compressed_size;
 335         struct page **pages;
 336         unsigned long nr_pages;
 337         int compress_type;
 338         struct list_head list;
 339 };
 340
 341 struct async_cow {
 342         struct inode *inode;
 343         struct btrfs_root *root;
 344         struct page *locked_page;
 345         u64 start;
 346         u64 end;
 347         struct list_head extents;
 348         struct btrfs_work work;
 349 };
 350
 351 static noinline int add_async_extent(struct async_cow *cow,
 352                                      u64 start, u64 ram_size,
 353                                      u64 compressed_size,
 354                                      struct page **pages,
 355                                      unsigned long nr_pages,
 356                                      int compress_type)
 357 {
 358         struct async_extent *async_extent;
 359
 360         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 361         BUG_ON(!async_extent); /* -ENOMEM */
 362         async_extent->start = start;
 363         async_extent->ram_size = ram_size;
 364         async_extent->compressed_size = compressed_size;
 365         async_extent->pages = pages;
 366         async_extent->nr_pages = nr_pages;
 367         async_extent->compress_type = compress_type;
 368         list_add_tail(&async_extent->list, &cow->extents);
 369         return 0;
 370 }
 371
 372 static inline int inode_need_compress(struct inode *inode)
 373 {
 374         struct btrfs_root *root = BTRFS_I(inode)->root;
 375
 376         /* force compress */
 377         if (btrfs_test_opt(root, FORCE_COMPRESS))
 378                 return 1;
 379         /* bad compression ratios */
 380         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 381                 return 0;
 382         if (btrfs_test_opt(root, COMPRESS) ||
 383             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 384             BTRFS_I(inode)->force_compress)
 385                 return 1;
 386         return 0;
 387 }
 388
 389 /*
 390  * we create compressed extents in two phases.  The first
 391  * phase compresses a range of pages that have already been
 392  * locked (both pages and state bits are locked).
 393  *
 394  * This is done inside an ordered work queue, and the compression
 395  * is spread across many cpus.  The actual IO submission is step
 396  * two, and the ordered work queue takes care of making sure that
 397  * happens in the same order things were put onto the queue by
 398  * writepages and friends.
 399  *
 400  * If this code finds it can't get good compression, it puts an
 401  * entry onto the work queue to write the uncompressed bytes.  This
 402  * makes sure that both compressed inodes and uncompressed inodes
 403  * are written in the same order that the flusher thread sent them
 404  * down.
 405  */
 406 static noinline void compress_file_range(struct inode *inode,
 407                                         struct page *locked_page,
 408                                         u64 start, u64 end,
 409                                         struct async_cow *async_cow,
 410                                         int *num_added)
 411 {
 412         struct btrfs_root *root = BTRFS_I(inode)->root;
 413         u64 num_bytes;
 414         u64 blocksize = root->sectorsize;
 415         u64 actual_end;
 416         u64 isize = i_size_read(inode);
 417         int ret = 0;
 418         struct page **pages = NULL;
 419         unsigned long nr_pages;
 420         unsigned long nr_pages_ret = 0;
 421         unsigned long total_compressed = 0;
 422         unsigned long total_in = 0;
 423         unsigned long max_compressed = SZ_128K;
 424         unsigned long max_uncompressed = SZ_128K;
 425         int i;
 426         int will_compress;
 427         int compress_type = root->fs_info->compress_type;
 428         int redirty = 0;
 429
 430         /* if this is a small write inside eof, kick off a defrag */
 431         if ((end - start + 1) < SZ_16K &&
 432             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 433                 btrfs_add_inode_defrag(NULL, inode);
 434
 435         actual_end = min_t(u64, isize, end + 1);
 436 again:
 437         will_compress = 0;
 438         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 439         nr_pages = min_t(unsigned long, nr_pages, SZ_128K / PAGE_SIZE);
 440
 441         /*
 442          * we don't want to send crud past the end of i_size through
 443          * compression, that's just a waste of CPU time.  So, if the
 444          * end of the file is before the start of our current
 445          * requested range of bytes, we bail out to the uncompressed
 446          * cleanup code that can deal with all of this.
 447          *
 448          * It isn't really the fastest way to fix things, but this is a
 449          * very uncommon corner.
 450          */
 451         if (actual_end <= start)
 452                 goto cleanup_and_bail_uncompressed;
 453
 454         total_compressed = actual_end - start;
 455
 456         /*
 457          * skip compression for a small file range(<=blocksize) that
 458          * isn't an inline extent, since it doesn't save disk space at all.
 459          */
 460         if (total_compressed <= blocksize &&
 461            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 462                 goto cleanup_and_bail_uncompressed;
 463
 464         /* we want to make sure that amount of ram required to uncompress
 465          * an extent is reasonable, so we limit the total size in ram
 466          * of a compressed extent to 128k.  This is a crucial number
 467          * because it also controls how easily we can spread reads across
 468          * cpus for decompression.
 469          *
 470          * We also want to make sure the amount of IO required to do
 471          * a random read is reasonably small, so we limit the size of
 472          * a compressed extent to 128k.
 473          */
 474         total_compressed = min(total_compressed, max_uncompressed);
 475         num_bytes = ALIGN(end - start + 1, blocksize);
 476         num_bytes = max(blocksize,  num_bytes);
 477         total_in = 0;
 478         ret = 0;
 479
 480         /*
 481          * we do compression for mount -o compress and when the
 482          * inode has not been flagged as nocompress.  This flag can
 483          * change at any time if we discover bad compression ratios.
 484          */
 485         if (inode_need_compress(inode)) {
 486                 WARN_ON(pages);
 487                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 488                 if (!pages) {
 489                         /* just bail out to the uncompressed code */
 490                         goto cont;
 491                 }
 492
 493                 if (BTRFS_I(inode)->force_compress)
 494                         compress_type = BTRFS_I(inode)->force_compress;
 495
 496                 /*
 497                  * we need to call clear_page_dirty_for_io on each
 498                  * page in the range.  Otherwise applications with the file
 499                  * mmap'd can wander in and change the page contents while
 500                  * we are compressing them.
 501                  *
 502                  * If the compression fails for any reason, we set the pages
 503                  * dirty again later on.
 504                  */
 505                 extent_range_clear_dirty_for_io(inode, start, end);
 506                 redirty = 1;
 507                 ret = btrfs_compress_pages(compress_type,
 508                                            inode->i_mapping, start,
 509                                            total_compressed, pages,
 510                                            nr_pages, &nr_pages_ret,
 511                                            &total_in,
 512                                            &total_compressed,
 513                                            max_compressed);
 514
 515                 if (!ret) {
 516                         unsigned long offset = total_compressed &
 517                                 (PAGE_SIZE - 1);
 518                         struct page *page = pages[nr_pages_ret - 1];
 519                         char *kaddr;
 520
 521                         /* zero the tail end of the last page, we might be
 522                          * sending it down to disk
 523                          */
 524                         if (offset) {
 525                                 kaddr = kmap_atomic(page);
 526                                 memset(kaddr + offset, 0,
 527                                        PAGE_SIZE - offset);
 528                                 kunmap_atomic(kaddr);
 529                         }
 530                         will_compress = 1;
 531                 }
 532         }
 533 cont:
 534         if (start == 0) {
 535                 /* lets try to make an inline extent */
 536                 if (ret || total_in < (actual_end - start)) {
 537                         /* we didn't compress the entire range, try
 538                          * to make an uncompressed inline extent.
 539                          */
 540                         ret = cow_file_range_inline(root, inode, start, end,
 541                                                     0, 0, NULL);
 542                 } else {
 543                         /* try making a compressed inline extent */
 544                         ret = cow_file_range_inline(root, inode, start, end,
 545                                                     total_compressed,
 546                                                     compress_type, pages);
 547                 }
 548                 if (ret <= 0) {
 549                         unsigned long clear_flags = EXTENT_DELALLOC |
 550                                 EXTENT_DEFRAG;
 551                         unsigned long page_error_op;
 552
 553                         clear_flags |= (ret < 0) ? EXTENT_DO_ACCOUNTING : 0;
 554                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 555
 556                         /*
 557                          * inline extent creation worked or returned error,
 558                          * we don't need to create any more async work items.
 559                          * Unlock and free up our temp pages.
 560                          */
 561                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 562                                                      clear_flags, PAGE_UNLOCK |
 563                                                      PAGE_CLEAR_DIRTY |
 564                                                      PAGE_SET_WRITEBACK |
 565                                                      page_error_op |
 566                                                      PAGE_END_WRITEBACK);
 567                         goto free_pages_out;
 568                 }
 569         }
 570
 571         if (will_compress) {
 572                 /*
 573                  * we aren't doing an inline extent round the compressed size
 574                  * up to a block size boundary so the allocator does sane
 575                  * things
 576                  */
 577                 total_compressed = ALIGN(total_compressed, blocksize);
 578
 579                 /*
 580                  * one last check to make sure the compression is really a
 581                  * win, compare the page count read with the blocks on disk
 582                  */
 583                 total_in = ALIGN(total_in, PAGE_SIZE);
 584                 if (total_compressed >= total_in) {
 585                         will_compress = 0;
 586                 } else {
 587                         num_bytes = total_in;
 588                 }
 589         }
 590         if (!will_compress && pages) {
 591                 /*
 592                  * the compression code ran but failed to make things smaller,
 593                  * free any pages it allocated and our page pointer array
 594                  */
 595                 for (i = 0; i < nr_pages_ret; i++) {
 596                         WARN_ON(pages[i]->mapping);
 597                         put_page(pages[i]);
 598                 }
 599                 kfree(pages);
 600                 pages = NULL;
 601                 total_compressed = 0;
 602                 nr_pages_ret = 0;
 603
 604                 /* flag the file so we don't compress in the future */
 605                 if (!btrfs_test_opt(root, FORCE_COMPRESS) &&
 606                     !(BTRFS_I(inode)->force_compress)) {
 607                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 608                 }
 609         }
 610         if (will_compress) {
 611                 *num_added += 1;
 612
 613                 /* the async work queues will take care of doing actual
 614                  * allocation on disk for these compressed pages,
 615                  * and will submit them to the elevator.
 616                  */
 617                 add_async_extent(async_cow, start, num_bytes,
 618                                  total_compressed, pages, nr_pages_ret,
 619                                  compress_type);
 620
 621                 if (start + num_bytes < end) {
 622                         start += num_bytes;
 623                         pages = NULL;
 624                         cond_resched();
 625                         goto again;
 626                 }
 627         } else {
 628 cleanup_and_bail_uncompressed:
 629                 /*
 630                  * No compression, but we still need to write the pages in
 631                  * the file we've been given so far.  redirty the locked
 632                  * page if it corresponds to our extent and set things up
 633                  * for the async work queue to run cow_file_range to do
 634                  * the normal delalloc dance
 635                  */
 636                 if (page_offset(locked_page) >= start &&
 637                     page_offset(locked_page) <= end) {
 638                         __set_page_dirty_nobuffers(locked_page);
 639                         /* unlocked later on in the async handlers */
 640                 }
 641                 if (redirty)
 642                         extent_range_redirty_for_io(inode, start, end);
 643                 add_async_extent(async_cow, start, end - start + 1,
 644                                  0, NULL, 0, BTRFS_COMPRESS_NONE);
 645                 *num_added += 1;
 646         }
 647
 648         return;
 649
 650 free_pages_out:
 651         for (i = 0; i < nr_pages_ret; i++) {
 652                 WARN_ON(pages[i]->mapping);
 653                 put_page(pages[i]);
 654         }
 655         kfree(pages);
 656 }
 657
 658 static void free_async_extent_pages(struct async_extent *async_extent)
 659 {
 660         int i;
 661
 662         if (!async_extent->pages)
 663                 return;
 664
 665         for (i = 0; i < async_extent->nr_pages; i++) {
 666                 WARN_ON(async_extent->pages[i]->mapping);
 667                 put_page(async_extent->pages[i]);
 668         }
 669         kfree(async_extent->pages);
 670         async_extent->nr_pages = 0;
 671         async_extent->pages = NULL;
 672 }
 673
 674 /*
 675  * phase two of compressed writeback.  This is the ordered portion
 676  * of the code, which only gets called in the order the work was
 677  * queued.  We walk all the async extents created by compress_file_range
 678  * and send them down to the disk.
 679  */
 680 static noinline void submit_compressed_extents(struct inode *inode,
 681                                               struct async_cow *async_cow)
 682 {
 683         struct async_extent *async_extent;
 684         u64 alloc_hint = 0;
 685         struct btrfs_key ins;
 686         struct extent_map *em;
 687         struct btrfs_root *root = BTRFS_I(inode)->root;
 688         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 689         struct extent_io_tree *io_tree;
 690         int ret = 0;
 691
 692 again:
 693         while (!list_empty(&async_cow->extents)) {
 694                 async_extent = list_entry(async_cow->extents.next,
 695                                           struct async_extent, list);
 696                 list_del(&async_extent->list);
 697
 698                 io_tree = &BTRFS_I(inode)->io_tree;
 699
 700 retry:
 701                 /* did the compression code fall back to uncompressed IO? */
 702                 if (!async_extent->pages) {
 703                         int page_started = 0;
 704                         unsigned long nr_written = 0;
 705
 706                         lock_extent(io_tree, async_extent->start,
 707                                          async_extent->start +
 708                                          async_extent->ram_size - 1);
 709
 710                         /* allocate blocks */
 711                         ret = cow_file_range(inode, async_cow->locked_page,
 712                                              async_extent->start,
 713                                              async_extent->start +
 714                                              async_extent->ram_size - 1,
 715                                              &page_started, &nr_written, 0);
 716
 717                         /* JDM XXX */
 718
 719                         /*
 720                          * if page_started, cow_file_range inserted an
 721                          * inline extent and took care of all the unlocking
 722                          * and IO for us.  Otherwise, we need to submit
 723                          * all those pages down to the drive.
 724                          */
 725                         if (!page_started && !ret)
 726                                 extent_write_locked_range(io_tree,
 727                                                   inode, async_extent->start,
 728                                                   async_extent->start +
 729                                                   async_extent->ram_size - 1,
 730                                                   btrfs_get_extent,
 731                                                   WB_SYNC_ALL);
 732                         else if (ret)
 733                                 unlock_page(async_cow->locked_page);
 734                         kfree(async_extent);
 735                         cond_resched();
 736                         continue;
 737                 }
 738
 739                 lock_extent(io_tree, async_extent->start,
 740                             async_extent->start + async_extent->ram_size - 1);
 741
 742                 ret = btrfs_reserve_extent(root,
 743                                            async_extent->compressed_size,
 744                                            async_extent->compressed_size,
 745                                            0, alloc_hint, &ins, 1, 1);
 746                 if (ret) {
 747                         free_async_extent_pages(async_extent);
 748
 749                         if (ret == -ENOSPC) {
 750                                 unlock_extent(io_tree, async_extent->start,
 751                                               async_extent->start +
 752                                               async_extent->ram_size - 1);
 753
 754                                 /*
 755                                  * we need to redirty the pages if we decide to
 756                                  * fallback to uncompressed IO, otherwise we
 757                                  * will not submit these pages down to lower
 758                                  * layers.
 759                                  */
 760                                 extent_range_redirty_for_io(inode,
 761                                                 async_extent->start,
 762                                                 async_extent->start +
 763                                                 async_extent->ram_size - 1);
 764
 765                                 goto retry;
 766                         }
 767                         goto out_free;
 768                 }
 769                 /*
 770                  * here we're doing allocation and writeback of the
 771                  * compressed pages
 772                  */
 773                 btrfs_drop_extent_cache(inode, async_extent->start,
 774                                         async_extent->start +
 775                                         async_extent->ram_size - 1, 0);
 776
 777                 em = alloc_extent_map();
 778                 if (!em) {
 779                         ret = -ENOMEM;
 780                         goto out_free_reserve;
 781                 }
 782                 em->start = async_extent->start;
 783                 em->len = async_extent->ram_size;
 784                 em->orig_start = em->start;
 785                 em->mod_start = em->start;
 786                 em->mod_len = em->len;
 787
 788                 em->block_start = ins.objectid;
 789                 em->block_len = ins.offset;
 790                 em->orig_block_len = ins.offset;
 791                 em->ram_bytes = async_extent->ram_size;
 792                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 793                 em->compress_type = async_extent->compress_type;
 794                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 795                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 796                 em->generation = -1;
 797
 798                 while (1) {
 799                         write_lock(&em_tree->lock);
 800                         ret = add_extent_mapping(em_tree, em, 1);
 801                         write_unlock(&em_tree->lock);
 802                         if (ret != -EEXIST) {
 803                                 free_extent_map(em);
 804                                 break;
 805                         }
 806                         btrfs_drop_extent_cache(inode, async_extent->start,
 807                                                 async_extent->start +
 808                                                 async_extent->ram_size - 1, 0);
 809                 }
 810
 811                 if (ret)
 812                         goto out_free_reserve;
 813
 814                 ret = btrfs_add_ordered_extent_compress(inode,
 815                                                 async_extent->start,
 816                                                 ins.objectid,
 817                                                 async_extent->ram_size,
 818                                                 ins.offset,
 819                                                 BTRFS_ORDERED_COMPRESSED,
 820                                                 async_extent->compress_type);
 821                 if (ret) {
 822                         btrfs_drop_extent_cache(inode, async_extent->start,
 823                                                 async_extent->start +
 824                                                 async_extent->ram_size - 1, 0);
 825                         goto out_free_reserve;
 826                 }
 827                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
 828
 829                 /*
 830                  * clear dirty, set writeback and unlock the pages.
 831                  */
 832                 extent_clear_unlock_delalloc(inode, async_extent->start,
 833                                 async_extent->start +
 834                                 async_extent->ram_size - 1,
 835                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 836                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 837                                 PAGE_SET_WRITEBACK);
 838                 ret = btrfs_submit_compressed_write(inode,
 839                                     async_extent->start,
 840                                     async_extent->ram_size,
 841                                     ins.objectid,
 842                                     ins.offset, async_extent->pages,
 843                                     async_extent->nr_pages);
 844                 if (ret) {
 845                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 846                         struct page *p = async_extent->pages[0];
 847                         const u64 start = async_extent->start;
 848                         const u64 end = start + async_extent->ram_size - 1;
 849
 850                         p->mapping = inode->i_mapping;
 851                         tree->ops->writepage_end_io_hook(p, start, end,
 852                                                          NULL, 0);
 853                         p->mapping = NULL;
 854                         extent_clear_unlock_delalloc(inode, start, end, NULL, 0,
 855                                                      PAGE_END_WRITEBACK |
 856                                                      PAGE_SET_ERROR);
 857                         free_async_extent_pages(async_extent);
 858                 }
 859                 alloc_hint = ins.objectid + ins.offset;
 860                 kfree(async_extent);
 861                 cond_resched();
 862         }
 863         return;
 864 out_free_reserve:
 865         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
 866         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
 867 out_free:
 868         extent_clear_unlock_delalloc(inode, async_extent->start,
 869                                      async_extent->start +
 870                                      async_extent->ram_size - 1,
 871                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 872                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 873                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 874                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 875                                      PAGE_SET_ERROR);
 876         free_async_extent_pages(async_extent);
 877         kfree(async_extent);
 878         goto again;
 879 }
 880
 881 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 882                                       u64 num_bytes)
 883 {
 884         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 885         struct extent_map *em;
 886         u64 alloc_hint = 0;
 887
 888         read_lock(&em_tree->lock);
 889         em = search_extent_mapping(em_tree, start, num_bytes);
 890         if (em) {
 891                 /*
 892                  * if block start isn't an actual block number then find the
 893                  * first block in this inode and use that as a hint.  If that
 894                  * block is also bogus then just don't worry about it.
 895                  */
 896                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 897                         free_extent_map(em);
 898                         em = search_extent_mapping(em_tree, 0, 0);
 899                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 900                                 alloc_hint = em->block_start;
 901                         if (em)
 902                                 free_extent_map(em);
 903                 } else {
 904                         alloc_hint = em->block_start;
 905                         free_extent_map(em);
 906                 }
 907         }
 908         read_unlock(&em_tree->lock);
 909
 910         return alloc_hint;
 911 }
 912
 913 /*
 914  * when extent_io.c finds a delayed allocation range in the file,
 915  * the call backs end up in this code.  The basic idea is to
 916  * allocate extents on disk for the range, and create ordered data structs
 917  * in ram to track those extents.
 918  *
 919  * locked_page is the page that writepage had locked already.  We use
 920  * it to make sure we don't do extra locks or unlocks.
 921  *
 922  * *page_started is set to one if we unlock locked_page and do everything
 923  * required to start IO on it.  It may be clean and already done with
 924  * IO when we return.
 925  */
 926 static noinline int cow_file_range(struct inode *inode,
 927                                    struct page *locked_page,
 928                                    u64 start, u64 end, int *page_started,
 929                                    unsigned long *nr_written,
 930                                    int unlock)
 931 {
 932         struct btrfs_root *root = BTRFS_I(inode)->root;
 933         u64 alloc_hint = 0;
 934         u64 num_bytes;
 935         unsigned long ram_size;
 936         u64 disk_num_bytes;
 937         u64 cur_alloc_size;
 938         u64 blocksize = root->sectorsize;
 939         struct btrfs_key ins;
 940         struct extent_map *em;
 941         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 942         int ret = 0;
 943
 944         if (btrfs_is_free_space_inode(inode)) {
 945                 WARN_ON_ONCE(1);
 946                 ret = -EINVAL;
 947                 goto out_unlock;
 948         }
 949
 950         num_bytes = ALIGN(end - start + 1, blocksize);
 951         num_bytes = max(blocksize,  num_bytes);
 952         disk_num_bytes = num_bytes;
 953
 954         /* if this is a small write inside eof, kick off defrag */
 955         if (num_bytes < SZ_64K &&
 956             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 957                 btrfs_add_inode_defrag(NULL, inode);
 958
 959         if (start == 0) {
 960                 /* lets try to make an inline extent */
 961                 ret = cow_file_range_inline(root, inode, start, end, 0, 0,
 962                                             NULL);
 963                 if (ret == 0) {
 964                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 965                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 966                                      EXTENT_DEFRAG, PAGE_UNLOCK |
 967                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 968                                      PAGE_END_WRITEBACK);
 969
 970                         *nr_written = *nr_written +
 971                              (end - start + PAGE_SIZE) / PAGE_SIZE;
 972                         *page_started = 1;
 973                         goto out;
 974                 } else if (ret < 0) {
 975                         goto out_unlock;
 976                 }
 977         }
 978
 979         BUG_ON(disk_num_bytes >
 980                btrfs_super_total_bytes(root->fs_info->super_copy));
 981
 982         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 983         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
 984
 985         while (disk_num_bytes > 0) {
 986                 unsigned long op;
 987
 988                 cur_alloc_size = disk_num_bytes;
 989                 ret = btrfs_reserve_extent(root, cur_alloc_size,
 990                                            root->sectorsize, 0, alloc_hint,
 991                                            &ins, 1, 1);
 992                 if (ret < 0)
 993                         goto out_unlock;
 994
 995                 em = alloc_extent_map();
 996                 if (!em) {
 997                         ret = -ENOMEM;
 998                         goto out_reserve;
 999                 }
1000                 em->start = start;
1001                 em->orig_start = em->start;
1002                 ram_size = ins.offset;
1003                 em->len = ins.offset;
1004                 em->mod_start = em->start;
1005                 em->mod_len = em->len;
1006
1007                 em->block_start = ins.objectid;
1008                 em->block_len = ins.offset;
1009                 em->orig_block_len = ins.offset;
1010                 em->ram_bytes = ram_size;
1011                 em->bdev = root->fs_info->fs_devices->latest_bdev;
1012                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
1013                 em->generation = -1;
1014
1015                 while (1) {
1016                         write_lock(&em_tree->lock);
1017                         ret = add_extent_mapping(em_tree, em, 1);
1018                         write_unlock(&em_tree->lock);
1019                         if (ret != -EEXIST) {
1020                                 free_extent_map(em);
1021                                 break;
1022                         }
1023                         btrfs_drop_extent_cache(inode, start,
1024                                                 start + ram_size - 1, 0);
1025                 }
1026                 if (ret)
1027                         goto out_reserve;
1028
1029                 cur_alloc_size = ins.offset;
1030                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1031                                                ram_size, cur_alloc_size, 0);
1032                 if (ret)
1033                         goto out_drop_extent_cache;
1034
1035                 if (root->root_key.objectid ==
1036                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1037                         ret = btrfs_reloc_clone_csums(inode, start,
1038                                                       cur_alloc_size);
1039                         if (ret)
1040                                 goto out_drop_extent_cache;
1041                 }
1042
1043                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
1044
1045                 if (disk_num_bytes < cur_alloc_size)
1046                         break;
1047
1048                 /* we're not doing compressed IO, don't unlock the first
1049                  * page (which the caller expects to stay locked), don't
1050                  * clear any dirty bits and don't set any writeback bits
1051                  *
1052                  * Do set the Private2 bit so we know this page was properly
1053                  * setup for writepage
1054                  */
1055                 op = unlock ? PAGE_UNLOCK : 0;
1056                 op |= PAGE_SET_PRIVATE2;
1057
1058                 extent_clear_unlock_delalloc(inode, start,
1059                                              start + ram_size - 1, locked_page,
1060                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1061                                              op);
1062                 disk_num_bytes -= cur_alloc_size;
1063                 num_bytes -= cur_alloc_size;
1064                 alloc_hint = ins.objectid + ins.offset;
1065                 start += cur_alloc_size;
1066         }
1067 out:
1068         return ret;
1069
1070 out_drop_extent_cache:
1071         btrfs_drop_extent_cache(inode, start, start + ram_size - 1, 0);
1072 out_reserve:
1073         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
1074         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
1075 out_unlock:
1076         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1077                                      EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
1078                                      EXTENT_DELALLOC | EXTENT_DEFRAG,
1079                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1080                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK);
1081         goto out;
1082 }
1083
1084 /*
1085  * work queue call back to started compression on a file and pages
1086  */
1087 static noinline void async_cow_start(struct btrfs_work *work)
1088 {
1089         struct async_cow *async_cow;
1090         int num_added = 0;
1091         async_cow = container_of(work, struct async_cow, work);
1092
1093         compress_file_range(async_cow->inode, async_cow->locked_page,
1094                             async_cow->start, async_cow->end, async_cow,
1095                             &num_added);
1096         if (num_added == 0) {
1097                 btrfs_add_delayed_iput(async_cow->inode);
1098                 async_cow->inode = NULL;
1099         }
1100 }
1101
1102 /*
1103  * work queue call back to submit previously compressed pages
1104  */
1105 static noinline void async_cow_submit(struct btrfs_work *work)
1106 {
1107         struct async_cow *async_cow;
1108         struct btrfs_root *root;
1109         unsigned long nr_pages;
1110
1111         async_cow = container_of(work, struct async_cow, work);
1112
1113         root = async_cow->root;
1114         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1115                 PAGE_SHIFT;
1116
1117         /*
1118          * atomic_sub_return implies a barrier for waitqueue_active
1119          */
1120         if (atomic_sub_return(nr_pages, &root->fs_info->async_delalloc_pages) <
1121             5 * SZ_1M &&
1122             waitqueue_active(&root->fs_info->async_submit_wait))
1123                 wake_up(&root->fs_info->async_submit_wait);
1124
1125         if (async_cow->inode)
1126                 submit_compressed_extents(async_cow->inode, async_cow);
1127 }
1128
1129 static noinline void async_cow_free(struct btrfs_work *work)
1130 {
1131         struct async_cow *async_cow;
1132         async_cow = container_of(work, struct async_cow, work);
1133         if (async_cow->inode)
1134                 btrfs_add_delayed_iput(async_cow->inode);
1135         kfree(async_cow);
1136 }
1137
1138 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1139                                 u64 start, u64 end, int *page_started,
1140                                 unsigned long *nr_written)
1141 {
1142         struct async_cow *async_cow;
1143         struct btrfs_root *root = BTRFS_I(inode)->root;
1144         unsigned long nr_pages;
1145         u64 cur_end;
1146         int limit = 10 * SZ_1M;
1147
1148         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1149                          1, 0, NULL, GFP_NOFS);
1150         while (start < end) {
1151                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1152                 BUG_ON(!async_cow); /* -ENOMEM */
1153                 async_cow->inode = igrab(inode);
1154                 async_cow->root = root;
1155                 async_cow->locked_page = locked_page;
1156                 async_cow->start = start;
1157
1158                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1159                     !btrfs_test_opt(root, FORCE_COMPRESS))
1160                         cur_end = end;
1161                 else
1162                         cur_end = min(end, start + SZ_512K - 1);
1163
1164                 async_cow->end = cur_end;
1165                 INIT_LIST_HEAD(&async_cow->extents);
1166
1167                 btrfs_init_work(&async_cow->work,
1168                                 btrfs_delalloc_helper,
1169                                 async_cow_start, async_cow_submit,
1170                                 async_cow_free);
1171
1172                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1173                         PAGE_SHIFT;
1174                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
1175
1176                 btrfs_queue_work(root->fs_info->delalloc_workers,
1177                                  &async_cow->work);
1178
1179                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
1180                         wait_event(root->fs_info->async_submit_wait,
1181                            (atomic_read(&root->fs_info->async_delalloc_pages) <
1182                             limit));
1183                 }
1184
1185                 while (atomic_read(&root->fs_info->async_submit_draining) &&
1186                       atomic_read(&root->fs_info->async_delalloc_pages)) {
1187                         wait_event(root->fs_info->async_submit_wait,
1188                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
1189                            0));
1190                 }
1191
1192                 *nr_written += nr_pages;
1193                 start = cur_end + 1;
1194         }
1195         *page_started = 1;
1196         return 0;
1197 }
1198
1199 static noinline int csum_exist_in_range(struct btrfs_root *root,
1200                                         u64 bytenr, u64 num_bytes)
1201 {
1202         int ret;
1203         struct btrfs_ordered_sum *sums;
1204         LIST_HEAD(list);
1205
1206         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
1207                                        bytenr + num_bytes - 1, &list, 0);
1208         if (ret == 0 && list_empty(&list))
1209                 return 0;
1210
1211         while (!list_empty(&list)) {
1212                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1213                 list_del(&sums->list);
1214                 kfree(sums);
1215         }
1216         return 1;
1217 }
1218
1219 /*
1220  * when nowcow writeback call back.  This checks for snapshots or COW copies
1221  * of the extents that exist in the file, and COWs the file as required.
1222  *
1223  * If no cow copies or snapshots exist, we write directly to the existing
1224  * blocks on disk
1225  */
1226 static noinline int run_delalloc_nocow(struct inode *inode,
1227                                        struct page *locked_page,
1228                               u64 start, u64 end, int *page_started, int force,
1229                               unsigned long *nr_written)
1230 {
1231         struct btrfs_root *root = BTRFS_I(inode)->root;
1232         struct btrfs_trans_handle *trans;
1233         struct extent_buffer *leaf;
1234         struct btrfs_path *path;
1235         struct btrfs_file_extent_item *fi;
1236         struct btrfs_key found_key;
1237         u64 cow_start;
1238         u64 cur_offset;
1239         u64 extent_end;
1240         u64 extent_offset;
1241         u64 disk_bytenr;
1242         u64 num_bytes;
1243         u64 disk_num_bytes;
1244         u64 ram_bytes;
1245         int extent_type;
1246         int ret, err;
1247         int type;
1248         int nocow;
1249         int check_prev = 1;
1250         bool nolock;
1251         u64 ino = btrfs_ino(inode);
1252
1253         path = btrfs_alloc_path();
1254         if (!path) {
1255                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1256                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1257                                              EXTENT_DO_ACCOUNTING |
1258                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1259                                              PAGE_CLEAR_DIRTY |
1260                                              PAGE_SET_WRITEBACK |
1261                                              PAGE_END_WRITEBACK);
1262                 return -ENOMEM;
1263         }
1264
1265         nolock = btrfs_is_free_space_inode(inode);
1266
1267         if (nolock)
1268                 trans = btrfs_join_transaction_nolock(root);
1269         else
1270                 trans = btrfs_join_transaction(root);
1271
1272         if (IS_ERR(trans)) {
1273                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1274                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1275                                              EXTENT_DO_ACCOUNTING |
1276                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1277                                              PAGE_CLEAR_DIRTY |
1278                                              PAGE_SET_WRITEBACK |
1279                                              PAGE_END_WRITEBACK);
1280                 btrfs_free_path(path);
1281                 return PTR_ERR(trans);
1282         }
1283
1284         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1285
1286         cow_start = (u64)-1;
1287         cur_offset = start;
1288         while (1) {
1289                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
1290                                                cur_offset, 0);
1291                 if (ret < 0)
1292                         goto error;
1293                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1294                         leaf = path->nodes[0];
1295                         btrfs_item_key_to_cpu(leaf, &found_key,
1296                                               path->slots[0] - 1);
1297                         if (found_key.objectid == ino &&
1298                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1299                                 path->slots[0]--;
1300                 }
1301                 check_prev = 0;
1302 next_slot:
1303                 leaf = path->nodes[0];
1304                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1305                         ret = btrfs_next_leaf(root, path);
1306                         if (ret < 0)
1307                                 goto error;
1308                         if (ret > 0)
1309                                 break;
1310                         leaf = path->nodes[0];
1311                 }
1312
1313                 nocow = 0;
1314                 disk_bytenr = 0;
1315                 num_bytes = 0;
1316                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1317
1318                 if (found_key.objectid > ino)
1319                         break;
1320                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1321                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1322                         path->slots[0]++;
1323                         goto next_slot;
1324                 }
1325                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1326                     found_key.offset > end)
1327                         break;
1328
1329                 if (found_key.offset > cur_offset) {
1330                         extent_end = found_key.offset;
1331                         extent_type = 0;
1332                         goto out_check;
1333                 }
1334
1335                 fi = btrfs_item_ptr(leaf, path->slots[0],
1336                                     struct btrfs_file_extent_item);
1337                 extent_type = btrfs_file_extent_type(leaf, fi);
1338
1339                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1340                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1341                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1342                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1343                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1344                         extent_end = found_key.offset +
1345                                 btrfs_file_extent_num_bytes(leaf, fi);
1346                         disk_num_bytes =
1347                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1348                         if (extent_end <= start) {
1349                                 path->slots[0]++;
1350                                 goto next_slot;
1351                         }
1352                         if (disk_bytenr == 0)
1353                                 goto out_check;
1354                         if (btrfs_file_extent_compression(leaf, fi) ||
1355                             btrfs_file_extent_encryption(leaf, fi) ||
1356                             btrfs_file_extent_other_encoding(leaf, fi))
1357                                 goto out_check;
1358                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1359                                 goto out_check;
1360                         if (btrfs_extent_readonly(root, disk_bytenr))
1361                                 goto out_check;
1362                         if (btrfs_cross_ref_exist(trans, root, ino,
1363                                                   found_key.offset -
1364                                                   extent_offset, disk_bytenr))
1365                                 goto out_check;
1366                         disk_bytenr += extent_offset;
1367                         disk_bytenr += cur_offset - found_key.offset;
1368                         num_bytes = min(end + 1, extent_end) - cur_offset;
1369                         /*
1370                          * if there are pending snapshots for this root,
1371                          * we fall into common COW way.
1372                          */
1373                         if (!nolock) {
1374                                 err = btrfs_start_write_no_snapshoting(root);
1375                                 if (!err)
1376                                         goto out_check;
1377                         }
1378                         /*
1379                          * force cow if csum exists in the range.
1380                          * this ensure that csum for a given extent are
1381                          * either valid or do not exist.
1382                          */
1383                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1384                                 goto out_check;
1385                         if (!btrfs_inc_nocow_writers(root->fs_info,
1386                                                      disk_bytenr))
1387                                 goto out_check;
1388                         nocow = 1;
1389                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1390                         extent_end = found_key.offset +
1391                                 btrfs_file_extent_inline_len(leaf,
1392                                                      path->slots[0], fi);
1393                         extent_end = ALIGN(extent_end, root->sectorsize);
1394                 } else {
1395                         BUG_ON(1);
1396                 }
1397 out_check:
1398                 if (extent_end <= start) {
1399                         path->slots[0]++;
1400                         if (!nolock && nocow)
1401                                 btrfs_end_write_no_snapshoting(root);
1402                         if (nocow)
1403                                 btrfs_dec_nocow_writers(root->fs_info,
1404                                                         disk_bytenr);
1405                         goto next_slot;
1406                 }
1407                 if (!nocow) {
1408                         if (cow_start == (u64)-1)
1409                                 cow_start = cur_offset;
1410                         cur_offset = extent_end;
1411                         if (cur_offset > end)
1412                                 break;
1413                         path->slots[0]++;
1414                         goto next_slot;
1415                 }
1416
1417                 btrfs_release_path(path);
1418                 if (cow_start != (u64)-1) {
1419                         ret = cow_file_range(inode, locked_page,
1420                                              cow_start, found_key.offset - 1,
1421                                              page_started, nr_written, 1);
1422                         if (ret) {
1423                                 if (!nolock && nocow)
1424                                         btrfs_end_write_no_snapshoting(root);
1425                                 if (nocow)
1426                                         btrfs_dec_nocow_writers(root->fs_info,
1427                                                                 disk_bytenr);
1428                                 goto error;
1429                         }
1430                         cow_start = (u64)-1;
1431                 }
1432
1433                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1434                         struct extent_map *em;
1435                         struct extent_map_tree *em_tree;
1436                         em_tree = &BTRFS_I(inode)->extent_tree;
1437                         em = alloc_extent_map();
1438                         BUG_ON(!em); /* -ENOMEM */
1439                         em->start = cur_offset;
1440                         em->orig_start = found_key.offset - extent_offset;
1441                         em->len = num_bytes;
1442                         em->block_len = num_bytes;
1443                         em->block_start = disk_bytenr;
1444                         em->orig_block_len = disk_num_bytes;
1445                         em->ram_bytes = ram_bytes;
1446                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1447                         em->mod_start = em->start;
1448                         em->mod_len = em->len;
1449                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1450                         set_bit(EXTENT_FLAG_FILLING, &em->flags);
1451                         em->generation = -1;
1452                         while (1) {
1453                                 write_lock(&em_tree->lock);
1454                                 ret = add_extent_mapping(em_tree, em, 1);
1455                                 write_unlock(&em_tree->lock);
1456                                 if (ret != -EEXIST) {
1457                                         free_extent_map(em);
1458                                         break;
1459                                 }
1460                                 btrfs_drop_extent_cache(inode, em->start,
1461                                                 em->start + em->len - 1, 0);
1462                         }
1463                         type = BTRFS_ORDERED_PREALLOC;
1464                 } else {
1465                         type = BTRFS_ORDERED_NOCOW;
1466                 }
1467
1468                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1469                                                num_bytes, num_bytes, type);
1470                 if (nocow)
1471                         btrfs_dec_nocow_writers(root->fs_info, disk_bytenr);
1472                 BUG_ON(ret); /* -ENOMEM */
1473
1474                 if (root->root_key.objectid ==
1475                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1476                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1477                                                       num_bytes);
1478                         if (ret) {
1479                                 if (!nolock && nocow)
1480                                         btrfs_end_write_no_snapshoting(root);
1481                                 goto error;
1482                         }
1483                 }
1484
1485                 extent_clear_unlock_delalloc(inode, cur_offset,
1486                                              cur_offset + num_bytes - 1,
1487                                              locked_page, EXTENT_LOCKED |
1488                                              EXTENT_DELALLOC, PAGE_UNLOCK |
1489                                              PAGE_SET_PRIVATE2);
1490                 if (!nolock && nocow)
1491                         btrfs_end_write_no_snapshoting(root);
1492                 cur_offset = extent_end;
1493                 if (cur_offset > end)
1494                         break;
1495         }
1496         btrfs_release_path(path);
1497
1498         if (cur_offset <= end && cow_start == (u64)-1) {
1499                 cow_start = cur_offset;
1500                 cur_offset = end;
1501         }
1502
1503         if (cow_start != (u64)-1) {
1504                 ret = cow_file_range(inode, locked_page, cow_start, end,
1505                                      page_started, nr_written, 1);
1506                 if (ret)
1507                         goto error;
1508         }
1509
1510 error:
1511         err = btrfs_end_transaction(trans, root);
1512         if (!ret)
1513                 ret = err;
1514
1515         if (ret && cur_offset < end)
1516                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1517                                              locked_page, EXTENT_LOCKED |
1518                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1519                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1520                                              PAGE_CLEAR_DIRTY |
1521                                              PAGE_SET_WRITEBACK |
1522                                              PAGE_END_WRITEBACK);
1523         btrfs_free_path(path);
1524         return ret;
1525 }
1526
1527 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1528 {
1529
1530         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1531             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1532                 return 0;
1533
1534         /*
1535          * @defrag_bytes is a hint value, no spinlock held here,
1536          * if is not zero, it means the file is defragging.
1537          * Force cow if given extent needs to be defragged.
1538          */
1539         if (BTRFS_I(inode)->defrag_bytes &&
1540             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1541                            EXTENT_DEFRAG, 0, NULL))
1542                 return 1;
1543
1544         return 0;
1545 }
1546
1547 /*
1548  * extent_io.c call back to do delayed allocation processing
1549  */
1550 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1551                               u64 start, u64 end, int *page_started,
1552                               unsigned long *nr_written)
1553 {
1554         int ret;
1555         int force_cow = need_force_cow(inode, start, end);
1556
1557         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1558                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1559                                          page_started, 1, nr_written);
1560         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1561                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1562                                          page_started, 0, nr_written);
1563         } else if (!inode_need_compress(inode)) {
1564                 ret = cow_file_range(inode, locked_page, start, end,
1565                                       page_started, nr_written, 1);
1566         } else {
1567                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1568                         &BTRFS_I(inode)->runtime_flags);
1569                 ret = cow_file_range_async(inode, locked_page, start, end,
1570                                            page_started, nr_written);
1571         }
1572         return ret;
1573 }
1574
1575 static void btrfs_split_extent_hook(struct inode *inode,
1576                                     struct extent_state *orig, u64 split)
1577 {
1578         u64 size;
1579
1580         /* not delalloc, ignore it */
1581         if (!(orig->state & EXTENT_DELALLOC))
1582                 return;
1583
1584         size = orig->end - orig->start + 1;
1585         if (size > BTRFS_MAX_EXTENT_SIZE) {
1586                 u64 num_extents;
1587                 u64 new_size;
1588
1589                 /*
1590                  * See the explanation in btrfs_merge_extent_hook, the same
1591                  * applies here, just in reverse.
1592                  */
1593                 new_size = orig->end - split + 1;
1594                 num_extents = div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1595                                         BTRFS_MAX_EXTENT_SIZE);
1596                 new_size = split - orig->start;
1597                 num_extents += div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1598                                         BTRFS_MAX_EXTENT_SIZE);
1599                 if (div64_u64(size + BTRFS_MAX_EXTENT_SIZE - 1,
1600                               BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1601                         return;
1602         }
1603
1604         spin_lock(&BTRFS_I(inode)->lock);
1605         BTRFS_I(inode)->outstanding_extents++;
1606         spin_unlock(&BTRFS_I(inode)->lock);
1607 }
1608
1609 /*
1610  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1611  * extents so we can keep track of new extents that are just merged onto old
1612  * extents, such as when we are doing sequential writes, so we can properly
1613  * account for the metadata space we'll need.
1614  */
1615 static void btrfs_merge_extent_hook(struct inode *inode,
1616                                     struct extent_state *new,
1617                                     struct extent_state *other)
1618 {
1619         u64 new_size, old_size;
1620         u64 num_extents;
1621
1622         /* not delalloc, ignore it */
1623         if (!(other->state & EXTENT_DELALLOC))
1624                 return;
1625
1626         if (new->start > other->start)
1627                 new_size = new->end - other->start + 1;
1628         else
1629                 new_size = other->end - new->start + 1;
1630
1631         /* we're not bigger than the max, unreserve the space and go */
1632         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1633                 spin_lock(&BTRFS_I(inode)->lock);
1634                 BTRFS_I(inode)->outstanding_extents--;
1635                 spin_unlock(&BTRFS_I(inode)->lock);
1636                 return;
1637         }
1638
1639         /*
1640          * We have to add up either side to figure out how many extents were
1641          * accounted for before we merged into one big extent.  If the number of
1642          * extents we accounted for is <= the amount we need for the new range
1643          * then we can return, otherwise drop.  Think of it like this
1644          *
1645          * [ 4k][MAX_SIZE]
1646          *
1647          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1648          * need 2 outstanding extents, on one side we have 1 and the other side
1649          * we have 1 so they are == and we can return.  But in this case
1650          *
1651          * [MAX_SIZE+4k][MAX_SIZE+4k]
1652          *
1653          * Each range on their own accounts for 2 extents, but merged together
1654          * they are only 3 extents worth of accounting, so we need to drop in
1655          * this case.
1656          */
1657         old_size = other->end - other->start + 1;
1658         num_extents = div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1659                                 BTRFS_MAX_EXTENT_SIZE);
1660         old_size = new->end - new->start + 1;
1661         num_extents += div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1662                                  BTRFS_MAX_EXTENT_SIZE);
1663
1664         if (div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1665                       BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1666                 return;
1667
1668         spin_lock(&BTRFS_I(inode)->lock);
1669         BTRFS_I(inode)->outstanding_extents--;
1670         spin_unlock(&BTRFS_I(inode)->lock);
1671 }
1672
1673 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1674                                       struct inode *inode)
1675 {
1676         spin_lock(&root->delalloc_lock);
1677         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1678                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1679                               &root->delalloc_inodes);
1680                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1681                         &BTRFS_I(inode)->runtime_flags);
1682                 root->nr_delalloc_inodes++;
1683                 if (root->nr_delalloc_inodes == 1) {
1684                         spin_lock(&root->fs_info->delalloc_root_lock);
1685                         BUG_ON(!list_empty(&root->delalloc_root));
1686                         list_add_tail(&root->delalloc_root,
1687                                       &root->fs_info->delalloc_roots);
1688                         spin_unlock(&root->fs_info->delalloc_root_lock);
1689                 }
1690         }
1691         spin_unlock(&root->delalloc_lock);
1692 }
1693
1694 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1695                                      struct inode *inode)
1696 {
1697         spin_lock(&root->delalloc_lock);
1698         if (!list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1699                 list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1700                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1701                           &BTRFS_I(inode)->runtime_flags);
1702                 root->nr_delalloc_inodes--;
1703                 if (!root->nr_delalloc_inodes) {
1704                         spin_lock(&root->fs_info->delalloc_root_lock);
1705                         BUG_ON(list_empty(&root->delalloc_root));
1706                         list_del_init(&root->delalloc_root);
1707                         spin_unlock(&root->fs_info->delalloc_root_lock);
1708                 }
1709         }
1710         spin_unlock(&root->delalloc_lock);
1711 }
1712
1713 /*
1714  * extent_io.c set_bit_hook, used to track delayed allocation
1715  * bytes in this file, and to maintain the list of inodes that
1716  * have pending delalloc work to be done.
1717  */
1718 static void btrfs_set_bit_hook(struct inode *inode,
1719                                struct extent_state *state, unsigned *bits)
1720 {
1721
1722         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1723                 WARN_ON(1);
1724         /*
1725          * set_bit and clear bit hooks normally require _irqsave/restore
1726          * but in this case, we are only testing for the DELALLOC
1727          * bit, which is only set or cleared with irqs on
1728          */
1729         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1730                 struct btrfs_root *root = BTRFS_I(inode)->root;
1731                 u64 len = state->end + 1 - state->start;
1732                 bool do_list = !btrfs_is_free_space_inode(inode);
1733
1734                 if (*bits & EXTENT_FIRST_DELALLOC) {
1735                         *bits &= ~EXTENT_FIRST_DELALLOC;
1736                 } else {
1737                         spin_lock(&BTRFS_I(inode)->lock);
1738                         BTRFS_I(inode)->outstanding_extents++;
1739                         spin_unlock(&BTRFS_I(inode)->lock);
1740                 }
1741
1742                 /* For sanity tests */
1743                 if (btrfs_test_is_dummy_root(root))
1744                         return;
1745
1746                 __percpu_counter_add(&root->fs_info->delalloc_bytes, len,
1747                                      root->fs_info->delalloc_batch);
1748                 spin_lock(&BTRFS_I(inode)->lock);
1749                 BTRFS_I(inode)->delalloc_bytes += len;
1750                 if (*bits & EXTENT_DEFRAG)
1751                         BTRFS_I(inode)->defrag_bytes += len;
1752                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1753                                          &BTRFS_I(inode)->runtime_flags))
1754                         btrfs_add_delalloc_inodes(root, inode);
1755                 spin_unlock(&BTRFS_I(inode)->lock);
1756         }
1757 }
1758
1759 /*
1760  * extent_io.c clear_bit_hook, see set_bit_hook for why
1761  */
1762 static void btrfs_clear_bit_hook(struct inode *inode,
1763                                  struct extent_state *state,
1764                                  unsigned *bits)
1765 {
1766         u64 len = state->end + 1 - state->start;
1767         u64 num_extents = div64_u64(len + BTRFS_MAX_EXTENT_SIZE -1,
1768                                     BTRFS_MAX_EXTENT_SIZE);
1769
1770         spin_lock(&BTRFS_I(inode)->lock);
1771         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG))
1772                 BTRFS_I(inode)->defrag_bytes -= len;
1773         spin_unlock(&BTRFS_I(inode)->lock);
1774
1775         /*
1776          * set_bit and clear bit hooks normally require _irqsave/restore
1777          * but in this case, we are only testing for the DELALLOC
1778          * bit, which is only set or cleared with irqs on
1779          */
1780         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1781                 struct btrfs_root *root = BTRFS_I(inode)->root;
1782                 bool do_list = !btrfs_is_free_space_inode(inode);
1783
1784                 if (*bits & EXTENT_FIRST_DELALLOC) {
1785                         *bits &= ~EXTENT_FIRST_DELALLOC;
1786                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1787                         spin_lock(&BTRFS_I(inode)->lock);
1788                         BTRFS_I(inode)->outstanding_extents -= num_extents;
1789                         spin_unlock(&BTRFS_I(inode)->lock);
1790                 }
1791
1792                 /*
1793                  * We don't reserve metadata space for space cache inodes so we
1794                  * don't need to call dellalloc_release_metadata if there is an
1795                  * error.
1796                  */
1797                 if (*bits & EXTENT_DO_ACCOUNTING &&
1798                     root != root->fs_info->tree_root)
1799                         btrfs_delalloc_release_metadata(inode, len);
1800
1801                 /* For sanity tests. */
1802                 if (btrfs_test_is_dummy_root(root))
1803                         return;
1804
1805                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1806                     && do_list && !(state->state & EXTENT_NORESERVE))
1807                         btrfs_free_reserved_data_space_noquota(inode,
1808                                         state->start, len);
1809
1810                 __percpu_counter_add(&root->fs_info->delalloc_bytes, -len,
1811                                      root->fs_info->delalloc_batch);
1812                 spin_lock(&BTRFS_I(inode)->lock);
1813                 BTRFS_I(inode)->delalloc_bytes -= len;
1814                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1815                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1816                              &BTRFS_I(inode)->runtime_flags))
1817                         btrfs_del_delalloc_inode(root, inode);
1818                 spin_unlock(&BTRFS_I(inode)->lock);
1819         }
1820 }
1821
1822 /*
1823  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1824  * we don't create bios that span stripes or chunks
1825  */
1826 int btrfs_merge_bio_hook(int rw, struct page *page, unsigned long offset,
1827                          size_t size, struct bio *bio,
1828                          unsigned long bio_flags)
1829 {
1830         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1831         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1832         u64 length = 0;
1833         u64 map_length;
1834         int ret;
1835
1836         if (bio_flags & EXTENT_BIO_COMPRESSED)
1837                 return 0;
1838
1839         length = bio->bi_iter.bi_size;
1840         map_length = length;
1841         ret = btrfs_map_block(root->fs_info, rw, logical,
1842                               &map_length, NULL, 0);
1843         /* Will always return 0 with map_multi == NULL */
1844         BUG_ON(ret < 0);
1845         if (map_length < length + size)
1846                 return 1;
1847         return 0;
1848 }
1849
1850 /*
1851  * in order to insert checksums into the metadata in large chunks,
1852  * we wait until bio submission time.   All the pages in the bio are
1853  * checksummed and sums are attached onto the ordered extent record.
1854  *
1855  * At IO completion time the cums attached on the ordered extent record
1856  * are inserted into the btree
1857  */
1858 static int __btrfs_submit_bio_start(struct inode *inode, int rw,
1859                                     struct bio *bio, int mirror_num,
1860                                     unsigned long bio_flags,
1861                                     u64 bio_offset)
1862 {
1863         struct btrfs_root *root = BTRFS_I(inode)->root;
1864         int ret = 0;
1865
1866         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1867         BUG_ON(ret); /* -ENOMEM */
1868         return 0;
1869 }
1870
1871 /*
1872  * in order to insert checksums into the metadata in large chunks,
1873  * we wait until bio submission time.   All the pages in the bio are
1874  * checksummed and sums are attached onto the ordered extent record.
1875  *
1876  * At IO completion time the cums attached on the ordered extent record
1877  * are inserted into the btree
1878  */
1879 static int __btrfs_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
1880                           int mirror_num, unsigned long bio_flags,
1881                           u64 bio_offset)
1882 {
1883         struct btrfs_root *root = BTRFS_I(inode)->root;
1884         int ret;
1885
1886         ret = btrfs_map_bio(root, rw, bio, mirror_num, 1);
1887         if (ret) {
1888                 bio->bi_error = ret;
1889                 bio_endio(bio);
1890         }
1891         return ret;
1892 }
1893
1894 /*
1895  * extent_io.c submission hook. This does the right thing for csum calculation
1896  * on write, or reading the csums from the tree before a read
1897  */
1898 static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
1899                           int mirror_num, unsigned long bio_flags,
1900                           u64 bio_offset)
1901 {
1902         struct btrfs_root *root = BTRFS_I(inode)->root;
1903         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1904         int ret = 0;
1905         int skip_sum;
1906         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1907
1908         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1909
1910         if (btrfs_is_free_space_inode(inode))
1911                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1912
1913         if (!(rw & REQ_WRITE)) {
1914                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, metadata);
1915                 if (ret)
1916                         goto out;
1917
1918                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1919                         ret = btrfs_submit_compressed_read(inode, bio,
1920                                                            mirror_num,
1921                                                            bio_flags);
1922                         goto out;
1923                 } else if (!skip_sum) {
1924                         ret = btrfs_lookup_bio_sums(root, inode, bio, NULL);
1925                         if (ret)
1926                                 goto out;
1927                 }
1928                 goto mapit;
1929         } else if (async && !skip_sum) {
1930                 /* csum items have already been cloned */
1931                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1932                         goto mapit;
1933                 /* we're doing a write, do the async checksumming */
1934                 ret = btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1935                                    inode, rw, bio, mirror_num,
1936                                    bio_flags, bio_offset,
1937                                    __btrfs_submit_bio_start,
1938                                    __btrfs_submit_bio_done);
1939                 goto out;
1940         } else if (!skip_sum) {
1941                 ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1942                 if (ret)
1943                         goto out;
1944         }
1945
1946 mapit:
1947         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
1948
1949 out:
1950         if (ret < 0) {
1951                 bio->bi_error = ret;
1952                 bio_endio(bio);
1953         }
1954         return ret;
1955 }
1956
1957 /*
1958  * given a list of ordered sums record them in the inode.  This happens
1959  * at IO completion time based on sums calculated at bio submission time.
1960  */
1961 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1962                              struct inode *inode, u64 file_offset,
1963                              struct list_head *list)
1964 {
1965         struct btrfs_ordered_sum *sum;
1966
1967         list_for_each_entry(sum, list, list) {
1968                 trans->adding_csums = 1;
1969                 btrfs_csum_file_blocks(trans,
1970                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1971                 trans->adding_csums = 0;
1972         }
1973         return 0;
1974 }
1975
1976 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1977                               struct extent_state **cached_state)
1978 {
1979         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
1980         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1981                                    cached_state);
1982 }
1983
1984 /* see btrfs_writepage_start_hook for details on why this is required */
1985 struct btrfs_writepage_fixup {
1986         struct page *page;
1987         struct btrfs_work work;
1988 };
1989
1990 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
1991 {
1992         struct btrfs_writepage_fixup *fixup;
1993         struct btrfs_ordered_extent *ordered;
1994         struct extent_state *cached_state = NULL;
1995         struct page *page;
1996         struct inode *inode;
1997         u64 page_start;
1998         u64 page_end;
1999         int ret;
2000
2001         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2002         page = fixup->page;
2003 again:
2004         lock_page(page);
2005         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2006                 ClearPageChecked(page);
2007                 goto out_page;
2008         }
2009
2010         inode = page->mapping->host;
2011         page_start = page_offset(page);
2012         page_end = page_offset(page) + PAGE_SIZE - 1;
2013
2014         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2015                          &cached_state);
2016
2017         /* already ordered? We're done */
2018         if (PagePrivate2(page))
2019                 goto out;
2020
2021         ordered = btrfs_lookup_ordered_range(inode, page_start,
2022                                         PAGE_SIZE);
2023         if (ordered) {
2024                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2025                                      page_end, &cached_state, GFP_NOFS);
2026                 unlock_page(page);
2027                 btrfs_start_ordered_extent(inode, ordered, 1);
2028                 btrfs_put_ordered_extent(ordered);
2029                 goto again;
2030         }
2031
2032         ret = btrfs_delalloc_reserve_space(inode, page_start,
2033                                            PAGE_SIZE);
2034         if (ret) {
2035                 mapping_set_error(page->mapping, ret);
2036                 end_extent_writepage(page, ret, page_start, page_end);
2037                 ClearPageChecked(page);
2038                 goto out;
2039          }
2040
2041         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
2042         ClearPageChecked(page);
2043         set_page_dirty(page);
2044 out:
2045         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2046                              &cached_state, GFP_NOFS);
2047 out_page:
2048         unlock_page(page);
2049         put_page(page);
2050         kfree(fixup);
2051 }
2052
2053 /*
2054  * There are a few paths in the higher layers of the kernel that directly
2055  * set the page dirty bit without asking the filesystem if it is a
2056  * good idea.  This causes problems because we want to make sure COW
2057  * properly happens and the data=ordered rules are followed.
2058  *
2059  * In our case any range that doesn't have the ORDERED bit set
2060  * hasn't been properly setup for IO.  We kick off an async process
2061  * to fix it up.  The async helper will wait for ordered extents, set
2062  * the delalloc bit and make it safe to write the page.
2063  */
2064 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2065 {
2066         struct inode *inode = page->mapping->host;
2067         struct btrfs_writepage_fixup *fixup;
2068         struct btrfs_root *root = BTRFS_I(inode)->root;
2069
2070         /* this page is properly in the ordered list */
2071         if (TestClearPagePrivate2(page))
2072                 return 0;
2073
2074         if (PageChecked(page))
2075                 return -EAGAIN;
2076
2077         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2078         if (!fixup)
2079                 return -EAGAIN;
2080
2081         SetPageChecked(page);
2082         get_page(page);
2083         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2084                         btrfs_writepage_fixup_worker, NULL, NULL);
2085         fixup->page = page;
2086         btrfs_queue_work(root->fs_info->fixup_workers, &fixup->work);
2087         return -EBUSY;
2088 }
2089
2090 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2091                                        struct inode *inode, u64 file_pos,
2092                                        u64 disk_bytenr, u64 disk_num_bytes,
2093                                        u64 num_bytes, u64 ram_bytes,
2094                                        u8 compression, u8 encryption,
2095                                        u16 other_encoding, int extent_type)
2096 {
2097         struct btrfs_root *root = BTRFS_I(inode)->root;
2098         struct btrfs_file_extent_item *fi;
2099         struct btrfs_path *path;
2100         struct extent_buffer *leaf;
2101         struct btrfs_key ins;
2102         int extent_inserted = 0;
2103         int ret;
2104
2105         path = btrfs_alloc_path();
2106         if (!path)
2107                 return -ENOMEM;
2108
2109         /*
2110          * we may be replacing one extent in the tree with another.
2111          * The new extent is pinned in the extent map, and we don't want
2112          * to drop it from the cache until it is completely in the btree.
2113          *
2114          * So, tell btrfs_drop_extents to leave this extent in the cache.
2115          * the caller is expected to unpin it and allow it to be merged
2116          * with the others.
2117          */
2118         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2119                                    file_pos + num_bytes, NULL, 0,
2120                                    1, sizeof(*fi), &extent_inserted);
2121         if (ret)
2122                 goto out;
2123
2124         if (!extent_inserted) {
2125                 ins.objectid = btrfs_ino(inode);
2126                 ins.offset = file_pos;
2127                 ins.type = BTRFS_EXTENT_DATA_KEY;
2128
2129                 path->leave_spinning = 1;
2130                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2131                                               sizeof(*fi));
2132                 if (ret)
2133                         goto out;
2134         }
2135         leaf = path->nodes[0];
2136         fi = btrfs_item_ptr(leaf, path->slots[0],
2137                             struct btrfs_file_extent_item);
2138         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2139         btrfs_set_file_extent_type(leaf, fi, extent_type);
2140         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2141         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2142         btrfs_set_file_extent_offset(leaf, fi, 0);
2143         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2144         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2145         btrfs_set_file_extent_compression(leaf, fi, compression);
2146         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2147         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2148
2149         btrfs_mark_buffer_dirty(leaf);
2150         btrfs_release_path(path);
2151
2152         inode_add_bytes(inode, num_bytes);
2153
2154         ins.objectid = disk_bytenr;
2155         ins.offset = disk_num_bytes;
2156         ins.type = BTRFS_EXTENT_ITEM_KEY;
2157         ret = btrfs_alloc_reserved_file_extent(trans, root,
2158                                         root->root_key.objectid,
2159                                         btrfs_ino(inode), file_pos,
2160                                         ram_bytes, &ins);
2161         /*
2162          * Release the reserved range from inode dirty range map, as it is
2163          * already moved into delayed_ref_head
2164          */
2165         btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2166 out:
2167         btrfs_free_path(path);
2168
2169         return ret;
2170 }
2171
2172 /* snapshot-aware defrag */
2173 struct sa_defrag_extent_backref {
2174         struct rb_node node;
2175         struct old_sa_defrag_extent *old;
2176         u64 root_id;
2177         u64 inum;
2178         u64 file_pos;
2179         u64 extent_offset;
2180         u64 num_bytes;
2181         u64 generation;
2182 };
2183
2184 struct old_sa_defrag_extent {
2185         struct list_head list;
2186         struct new_sa_defrag_extent *new;
2187
2188         u64 extent_offset;
2189         u64 bytenr;
2190         u64 offset;
2191         u64 len;
2192         int count;
2193 };
2194
2195 struct new_sa_defrag_extent {
2196         struct rb_root root;
2197         struct list_head head;
2198         struct btrfs_path *path;
2199         struct inode *inode;
2200         u64 file_pos;
2201         u64 len;
2202         u64 bytenr;
2203         u64 disk_len;
2204         u8 compress_type;
2205 };
2206
2207 static int backref_comp(struct sa_defrag_extent_backref *b1,
2208                         struct sa_defrag_extent_backref *b2)
2209 {
2210         if (b1->root_id < b2->root_id)
2211                 return -1;
2212         else if (b1->root_id > b2->root_id)
2213                 return 1;
2214
2215         if (b1->inum < b2->inum)
2216                 return -1;
2217         else if (b1->inum > b2->inum)
2218                 return 1;
2219
2220         if (b1->file_pos < b2->file_pos)
2221                 return -1;
2222         else if (b1->file_pos > b2->file_pos)
2223                 return 1;
2224
2225         /*
2226          * [------------------------------] ===> (a range of space)
2227          *     |<--->|   |<---->| =============> (fs/file tree A)
2228          * |<---------------------------->| ===> (fs/file tree B)
2229          *
2230          * A range of space can refer to two file extents in one tree while
2231          * refer to only one file extent in another tree.
2232          *
2233          * So we may process a disk offset more than one time(two extents in A)
2234          * and locate at the same extent(one extent in B), then insert two same
2235          * backrefs(both refer to the extent in B).
2236          */
2237         return 0;
2238 }
2239
2240 static void backref_insert(struct rb_root *root,
2241                            struct sa_defrag_extent_backref *backref)
2242 {
2243         struct rb_node **p = &root->rb_node;
2244         struct rb_node *parent = NULL;
2245         struct sa_defrag_extent_backref *entry;
2246         int ret;
2247
2248         while (*p) {
2249                 parent = *p;
2250                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2251
2252                 ret = backref_comp(backref, entry);
2253                 if (ret < 0)
2254                         p = &(*p)->rb_left;
2255                 else
2256                         p = &(*p)->rb_right;
2257         }
2258
2259         rb_link_node(&backref->node, parent, p);
2260         rb_insert_color(&backref->node, root);
2261 }
2262
2263 /*
2264  * Note the backref might has changed, and in this case we just return 0.
2265  */
2266 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2267                                        void *ctx)
2268 {
2269         struct btrfs_file_extent_item *extent;
2270         struct btrfs_fs_info *fs_info;
2271         struct old_sa_defrag_extent *old = ctx;
2272         struct new_sa_defrag_extent *new = old->new;
2273         struct btrfs_path *path = new->path;
2274         struct btrfs_key key;
2275         struct btrfs_root *root;
2276         struct sa_defrag_extent_backref *backref;
2277         struct extent_buffer *leaf;
2278         struct inode *inode = new->inode;
2279         int slot;
2280         int ret;
2281         u64 extent_offset;
2282         u64 num_bytes;
2283
2284         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2285             inum == btrfs_ino(inode))
2286                 return 0;
2287
2288         key.objectid = root_id;
2289         key.type = BTRFS_ROOT_ITEM_KEY;
2290         key.offset = (u64)-1;
2291
2292         fs_info = BTRFS_I(inode)->root->fs_info;
2293         root = btrfs_read_fs_root_no_name(fs_info, &key);
2294         if (IS_ERR(root)) {
2295                 if (PTR_ERR(root) == -ENOENT)
2296                         return 0;
2297                 WARN_ON(1);
2298                 pr_debug("inum=%llu, offset=%llu, root_id=%llu\n",
2299                          inum, offset, root_id);
2300                 return PTR_ERR(root);
2301         }
2302
2303         key.objectid = inum;
2304         key.type = BTRFS_EXTENT_DATA_KEY;
2305         if (offset > (u64)-1 << 32)
2306                 key.offset = 0;
2307         else
2308                 key.offset = offset;
2309
2310         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2311         if (WARN_ON(ret < 0))
2312                 return ret;
2313         ret = 0;
2314
2315         while (1) {
2316                 cond_resched();
2317
2318                 leaf = path->nodes[0];
2319                 slot = path->slots[0];
2320
2321                 if (slot >= btrfs_header_nritems(leaf)) {
2322                         ret = btrfs_next_leaf(root, path);
2323                         if (ret < 0) {
2324                                 goto out;
2325                         } else if (ret > 0) {
2326                                 ret = 0;
2327                                 goto out;
2328                         }
2329                         continue;
2330                 }
2331
2332                 path->slots[0]++;
2333
2334                 btrfs_item_key_to_cpu(leaf, &key, slot);
2335
2336                 if (key.objectid > inum)
2337                         goto out;
2338
2339                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2340                         continue;
2341
2342                 extent = btrfs_item_ptr(leaf, slot,
2343                                         struct btrfs_file_extent_item);
2344
2345                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2346                         continue;
2347
2348                 /*
2349                  * 'offset' refers to the exact key.offset,
2350                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2351                  * (key.offset - extent_offset).
2352                  */
2353                 if (key.offset != offset)
2354                         continue;
2355
2356                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2357                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2358
2359                 if (extent_offset >= old->extent_offset + old->offset +
2360                     old->len || extent_offset + num_bytes <=
2361                     old->extent_offset + old->offset)
2362                         continue;
2363                 break;
2364         }
2365
2366         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2367         if (!backref) {
2368                 ret = -ENOENT;
2369                 goto out;
2370         }
2371
2372         backref->root_id = root_id;
2373         backref->inum = inum;
2374         backref->file_pos = offset;
2375         backref->num_bytes = num_bytes;
2376         backref->extent_offset = extent_offset;
2377         backref->generation = btrfs_file_extent_generation(leaf, extent);
2378         backref->old = old;
2379         backref_insert(&new->root, backref);
2380         old->count++;
2381 out:
2382         btrfs_release_path(path);
2383         WARN_ON(ret);
2384         return ret;
2385 }
2386
2387 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2388                                    struct new_sa_defrag_extent *new)
2389 {
2390         struct btrfs_fs_info *fs_info = BTRFS_I(new->inode)->root->fs_info;
2391         struct old_sa_defrag_extent *old, *tmp;
2392         int ret;
2393
2394         new->path = path;
2395
2396         list_for_each_entry_safe(old, tmp, &new->head, list) {
2397                 ret = iterate_inodes_from_logical(old->bytenr +
2398                                                   old->extent_offset, fs_info,
2399                                                   path, record_one_backref,
2400                                                   old);
2401                 if (ret < 0 && ret != -ENOENT)
2402                         return false;
2403
2404                 /* no backref to be processed for this extent */
2405                 if (!old->count) {
2406                         list_del(&old->list);
2407                         kfree(old);
2408                 }
2409         }
2410
2411         if (list_empty(&new->head))
2412                 return false;
2413
2414         return true;
2415 }
2416
2417 static int relink_is_mergable(struct extent_buffer *leaf,
2418                               struct btrfs_file_extent_item *fi,
2419                               struct new_sa_defrag_extent *new)
2420 {
2421         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2422                 return 0;
2423
2424         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2425                 return 0;
2426
2427         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2428                 return 0;
2429
2430         if (btrfs_file_extent_encryption(leaf, fi) ||
2431             btrfs_file_extent_other_encoding(leaf, fi))
2432                 return 0;
2433
2434         return 1;
2435 }
2436
2437 /*
2438  * Note the backref might has changed, and in this case we just return 0.
2439  */
2440 static noinline int relink_extent_backref(struct btrfs_path *path,
2441                                  struct sa_defrag_extent_backref *prev,
2442                                  struct sa_defrag_extent_backref *backref)
2443 {
2444         struct btrfs_file_extent_item *extent;
2445         struct btrfs_file_extent_item *item;
2446         struct btrfs_ordered_extent *ordered;
2447         struct btrfs_trans_handle *trans;
2448         struct btrfs_fs_info *fs_info;
2449         struct btrfs_root *root;
2450         struct btrfs_key key;
2451         struct extent_buffer *leaf;
2452         struct old_sa_defrag_extent *old = backref->old;
2453         struct new_sa_defrag_extent *new = old->new;
2454         struct inode *src_inode = new->inode;
2455         struct inode *inode;
2456         struct extent_state *cached = NULL;
2457         int ret = 0;
2458         u64 start;
2459         u64 len;
2460         u64 lock_start;
2461         u64 lock_end;
2462         bool merge = false;
2463         int index;
2464
2465         if (prev && prev->root_id == backref->root_id &&
2466             prev->inum == backref->inum &&
2467             prev->file_pos + prev->num_bytes == backref->file_pos)
2468                 merge = true;
2469
2470         /* step 1: get root */
2471         key.objectid = backref->root_id;
2472         key.type = BTRFS_ROOT_ITEM_KEY;
2473         key.offset = (u64)-1;
2474
2475         fs_info = BTRFS_I(src_inode)->root->fs_info;
2476         index = srcu_read_lock(&fs_info->subvol_srcu);
2477
2478         root = btrfs_read_fs_root_no_name(fs_info, &key);
2479         if (IS_ERR(root)) {
2480                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2481                 if (PTR_ERR(root) == -ENOENT)
2482                         return 0;
2483                 return PTR_ERR(root);
2484         }
2485
2486         if (btrfs_root_readonly(root)) {
2487                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2488                 return 0;
2489         }
2490
2491         /* step 2: get inode */
2492         key.objectid = backref->inum;
2493         key.type = BTRFS_INODE_ITEM_KEY;
2494         key.offset = 0;
2495
2496         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2497         if (IS_ERR(inode)) {
2498                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2499                 return 0;
2500         }
2501
2502         srcu_read_unlock(&fs_info->subvol_srcu, index);
2503
2504         /* step 3: relink backref */
2505         lock_start = backref->file_pos;
2506         lock_end = backref->file_pos + backref->num_bytes - 1;
2507         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2508                          &cached);
2509
2510         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2511         if (ordered) {
2512                 btrfs_put_ordered_extent(ordered);
2513                 goto out_unlock;
2514         }
2515
2516         trans = btrfs_join_transaction(root);
2517         if (IS_ERR(trans)) {
2518                 ret = PTR_ERR(trans);
2519                 goto out_unlock;
2520         }
2521
2522         key.objectid = backref->inum;
2523         key.type = BTRFS_EXTENT_DATA_KEY;
2524         key.offset = backref->file_pos;
2525
2526         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2527         if (ret < 0) {
2528                 goto out_free_path;
2529         } else if (ret > 0) {
2530                 ret = 0;
2531                 goto out_free_path;
2532         }
2533
2534         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2535                                 struct btrfs_file_extent_item);
2536
2537         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2538             backref->generation)
2539                 goto out_free_path;
2540
2541         btrfs_release_path(path);
2542
2543         start = backref->file_pos;
2544         if (backref->extent_offset < old->extent_offset + old->offset)
2545                 start += old->extent_offset + old->offset -
2546                          backref->extent_offset;
2547
2548         len = min(backref->extent_offset + backref->num_bytes,
2549                   old->extent_offset + old->offset + old->len);
2550         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2551
2552         ret = btrfs_drop_extents(trans, root, inode, start,
2553                                  start + len, 1);
2554         if (ret)
2555                 goto out_free_path;
2556 again:
2557         key.objectid = btrfs_ino(inode);
2558         key.type = BTRFS_EXTENT_DATA_KEY;
2559         key.offset = start;
2560
2561         path->leave_spinning = 1;
2562         if (merge) {
2563                 struct btrfs_file_extent_item *fi;
2564                 u64 extent_len;
2565                 struct btrfs_key found_key;
2566
2567                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2568                 if (ret < 0)
2569                         goto out_free_path;
2570
2571                 path->slots[0]--;
2572                 leaf = path->nodes[0];
2573                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2574
2575                 fi = btrfs_item_ptr(leaf, path->slots[0],
2576                                     struct btrfs_file_extent_item);
2577                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2578
2579                 if (extent_len + found_key.offset == start &&
2580                     relink_is_mergable(leaf, fi, new)) {
2581                         btrfs_set_file_extent_num_bytes(leaf, fi,
2582                                                         extent_len + len);
2583                         btrfs_mark_buffer_dirty(leaf);
2584                         inode_add_bytes(inode, len);
2585
2586                         ret = 1;
2587                         goto out_free_path;
2588                 } else {
2589                         merge = false;
2590                         btrfs_release_path(path);
2591                         goto again;
2592                 }
2593         }
2594
2595         ret = btrfs_insert_empty_item(trans, root, path, &key,
2596                                         sizeof(*extent));
2597         if (ret) {
2598                 btrfs_abort_transaction(trans, root, ret);
2599                 goto out_free_path;
2600         }
2601
2602         leaf = path->nodes[0];
2603         item = btrfs_item_ptr(leaf, path->slots[0],
2604                                 struct btrfs_file_extent_item);
2605         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2606         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2607         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2608         btrfs_set_file_extent_num_bytes(leaf, item, len);
2609         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2610         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2611         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2612         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2613         btrfs_set_file_extent_encryption(leaf, item, 0);
2614         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2615
2616         btrfs_mark_buffer_dirty(leaf);
2617         inode_add_bytes(inode, len);
2618         btrfs_release_path(path);
2619
2620         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2621                         new->disk_len, 0,
2622                         backref->root_id, backref->inum,
2623                         new->file_pos); /* start - extent_offset */
2624         if (ret) {
2625                 btrfs_abort_transaction(trans, root, ret);
2626                 goto out_free_path;
2627         }
2628
2629         ret = 1;
2630 out_free_path:
2631         btrfs_release_path(path);
2632         path->leave_spinning = 0;
2633         btrfs_end_transaction(trans, root);
2634 out_unlock:
2635         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2636                              &cached, GFP_NOFS);
2637         iput(inode);
2638         return ret;
2639 }
2640
2641 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2642 {
2643         struct old_sa_defrag_extent *old, *tmp;
2644
2645         if (!new)
2646                 return;
2647
2648         list_for_each_entry_safe(old, tmp, &new->head, list) {
2649                 kfree(old);
2650         }
2651         kfree(new);
2652 }
2653
2654 static void relink_file_extents(struct new_sa_defrag_extent *new)
2655 {
2656         struct btrfs_path *path;
2657         struct sa_defrag_extent_backref *backref;
2658         struct sa_defrag_extent_backref *prev = NULL;
2659         struct inode *inode;
2660         struct btrfs_root *root;
2661         struct rb_node *node;
2662         int ret;
2663
2664         inode = new->inode;
2665         root = BTRFS_I(inode)->root;
2666
2667         path = btrfs_alloc_path();
2668         if (!path)
2669                 return;
2670
2671         if (!record_extent_backrefs(path, new)) {
2672                 btrfs_free_path(path);
2673                 goto out;
2674         }
2675         btrfs_release_path(path);
2676
2677         while (1) {
2678                 node = rb_first(&new->root);
2679                 if (!node)
2680                         break;
2681                 rb_erase(node, &new->root);
2682
2683                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2684
2685                 ret = relink_extent_backref(path, prev, backref);
2686                 WARN_ON(ret < 0);
2687
2688                 kfree(prev);
2689
2690                 if (ret == 1)
2691                         prev = backref;
2692                 else
2693                         prev = NULL;
2694                 cond_resched();
2695         }
2696         kfree(prev);
2697
2698         btrfs_free_path(path);
2699 out:
2700         free_sa_defrag_extent(new);
2701
2702         atomic_dec(&root->fs_info->defrag_running);
2703         wake_up(&root->fs_info->transaction_wait);
2704 }
2705
2706 static struct new_sa_defrag_extent *
2707 record_old_file_extents(struct inode *inode,
2708                         struct btrfs_ordered_extent *ordered)
2709 {
2710         struct btrfs_root *root = BTRFS_I(inode)->root;
2711         struct btrfs_path *path;
2712         struct btrfs_key key;
2713         struct old_sa_defrag_extent *old;
2714         struct new_sa_defrag_extent *new;
2715         int ret;
2716
2717         new = kmalloc(sizeof(*new), GFP_NOFS);
2718         if (!new)
2719                 return NULL;
2720
2721         new->inode = inode;
2722         new->file_pos = ordered->file_offset;
2723         new->len = ordered->len;
2724         new->bytenr = ordered->start;
2725         new->disk_len = ordered->disk_len;
2726         new->compress_type = ordered->compress_type;
2727         new->root = RB_ROOT;
2728         INIT_LIST_HEAD(&new->head);
2729
2730         path = btrfs_alloc_path();
2731         if (!path)
2732                 goto out_kfree;
2733
2734         key.objectid = btrfs_ino(inode);
2735         key.type = BTRFS_EXTENT_DATA_KEY;
2736         key.offset = new->file_pos;
2737
2738         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2739         if (ret < 0)
2740                 goto out_free_path;
2741         if (ret > 0 && path->slots[0] > 0)
2742                 path->slots[0]--;
2743
2744         /* find out all the old extents for the file range */
2745         while (1) {
2746                 struct btrfs_file_extent_item *extent;
2747                 struct extent_buffer *l;
2748                 int slot;
2749                 u64 num_bytes;
2750                 u64 offset;
2751                 u64 end;
2752                 u64 disk_bytenr;
2753                 u64 extent_offset;
2754
2755                 l = path->nodes[0];
2756                 slot = path->slots[0];
2757
2758                 if (slot >= btrfs_header_nritems(l)) {
2759                         ret = btrfs_next_leaf(root, path);
2760                         if (ret < 0)
2761                                 goto out_free_path;
2762                         else if (ret > 0)
2763                                 break;
2764                         continue;
2765                 }
2766
2767                 btrfs_item_key_to_cpu(l, &key, slot);
2768
2769                 if (key.objectid != btrfs_ino(inode))
2770                         break;
2771                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2772                         break;
2773                 if (key.offset >= new->file_pos + new->len)
2774                         break;
2775
2776                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2777
2778                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2779                 if (key.offset + num_bytes < new->file_pos)
2780                         goto next;
2781
2782                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2783                 if (!disk_bytenr)
2784                         goto next;
2785
2786                 extent_offset = btrfs_file_extent_offset(l, extent);
2787
2788                 old = kmalloc(sizeof(*old), GFP_NOFS);
2789                 if (!old)
2790                         goto out_free_path;
2791
2792                 offset = max(new->file_pos, key.offset);
2793                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2794
2795                 old->bytenr = disk_bytenr;
2796                 old->extent_offset = extent_offset;
2797                 old->offset = offset - key.offset;
2798                 old->len = end - offset;
2799                 old->new = new;
2800                 old->count = 0;
2801                 list_add_tail(&old->list, &new->head);
2802 next:
2803                 path->slots[0]++;
2804                 cond_resched();
2805         }
2806
2807         btrfs_free_path(path);
2808         atomic_inc(&root->fs_info->defrag_running);
2809
2810         return new;
2811
2812 out_free_path:
2813         btrfs_free_path(path);
2814 out_kfree:
2815         free_sa_defrag_extent(new);
2816         return NULL;
2817 }
2818
2819 static void btrfs_release_delalloc_bytes(struct btrfs_root *root,
2820                                          u64 start, u64 len)
2821 {
2822         struct btrfs_block_group_cache *cache;
2823
2824         cache = btrfs_lookup_block_group(root->fs_info, start);
2825         ASSERT(cache);
2826
2827         spin_lock(&cache->lock);
2828         cache->delalloc_bytes -= len;
2829         spin_unlock(&cache->lock);
2830
2831         btrfs_put_block_group(cache);
2832 }
2833
2834 /* as ordered data IO finishes, this gets called so we can finish
2835  * an ordered extent if the range of bytes in the file it covers are
2836  * fully written.
2837  */
2838 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2839 {
2840         struct inode *inode = ordered_extent->inode;
2841         struct btrfs_root *root = BTRFS_I(inode)->root;
2842         struct btrfs_trans_handle *trans = NULL;
2843         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2844         struct extent_state *cached_state = NULL;
2845         struct new_sa_defrag_extent *new = NULL;
2846         int compress_type = 0;
2847         int ret = 0;
2848         u64 logical_len = ordered_extent->len;
2849         bool nolock;
2850         bool truncated = false;
2851
2852         nolock = btrfs_is_free_space_inode(inode);
2853
2854         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2855                 ret = -EIO;
2856                 goto out;
2857         }
2858
2859         btrfs_free_io_failure_record(inode, ordered_extent->file_offset,
2860                                      ordered_extent->file_offset +
2861                                      ordered_extent->len - 1);
2862
2863         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2864                 truncated = true;
2865                 logical_len = ordered_extent->truncated_len;
2866                 /* Truncated the entire extent, don't bother adding */
2867                 if (!logical_len)
2868                         goto out;
2869         }
2870
2871         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2872                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2873
2874                 /*
2875                  * For mwrite(mmap + memset to write) case, we still reserve
2876                  * space for NOCOW range.
2877                  * As NOCOW won't cause a new delayed ref, just free the space
2878                  */
2879                 btrfs_qgroup_free_data(inode, ordered_extent->file_offset,
2880                                        ordered_extent->len);
2881                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2882                 if (nolock)
2883                         trans = btrfs_join_transaction_nolock(root);
2884                 else
2885                         trans = btrfs_join_transaction(root);
2886                 if (IS_ERR(trans)) {
2887                         ret = PTR_ERR(trans);
2888                         trans = NULL;
2889                         goto out;
2890                 }
2891                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2892                 ret = btrfs_update_inode_fallback(trans, root, inode);
2893                 if (ret) /* -ENOMEM or corruption */
2894                         btrfs_abort_transaction(trans, root, ret);
2895                 goto out;
2896         }
2897
2898         lock_extent_bits(io_tree, ordered_extent->file_offset,
2899                          ordered_extent->file_offset + ordered_extent->len - 1,
2900                          &cached_state);
2901
2902         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2903                         ordered_extent->file_offset + ordered_extent->len - 1,
2904                         EXTENT_DEFRAG, 1, cached_state);
2905         if (ret) {
2906                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2907                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2908                         /* the inode is shared */
2909                         new = record_old_file_extents(inode, ordered_extent);
2910
2911                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2912                         ordered_extent->file_offset + ordered_extent->len - 1,
2913                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2914         }
2915
2916         if (nolock)
2917                 trans = btrfs_join_transaction_nolock(root);
2918         else
2919                 trans = btrfs_join_transaction(root);
2920         if (IS_ERR(trans)) {
2921                 ret = PTR_ERR(trans);
2922                 trans = NULL;
2923                 goto out_unlock;
2924         }
2925
2926         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2927
2928         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2929                 compress_type = ordered_extent->compress_type;
2930         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2931                 BUG_ON(compress_type);
2932                 ret = btrfs_mark_extent_written(trans, inode,
2933                                                 ordered_extent->file_offset,
2934                                                 ordered_extent->file_offset +
2935                                                 logical_len);
2936         } else {
2937                 BUG_ON(root == root->fs_info->tree_root);
2938                 ret = insert_reserved_file_extent(trans, inode,
2939                                                 ordered_extent->file_offset,
2940                                                 ordered_extent->start,
2941                                                 ordered_extent->disk_len,
2942                                                 logical_len, logical_len,
2943                                                 compress_type, 0, 0,
2944                                                 BTRFS_FILE_EXTENT_REG);
2945                 if (!ret)
2946                         btrfs_release_delalloc_bytes(root,
2947                                                      ordered_extent->start,
2948                                                      ordered_extent->disk_len);
2949         }
2950         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2951                            ordered_extent->file_offset, ordered_extent->len,
2952                            trans->transid);
2953         if (ret < 0) {
2954                 btrfs_abort_transaction(trans, root, ret);
2955                 goto out_unlock;
2956         }
2957
2958         add_pending_csums(trans, inode, ordered_extent->file_offset,
2959                           &ordered_extent->list);
2960
2961         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2962         ret = btrfs_update_inode_fallback(trans, root, inode);
2963         if (ret) { /* -ENOMEM or corruption */
2964                 btrfs_abort_transaction(trans, root, ret);
2965                 goto out_unlock;
2966         }
2967         ret = 0;
2968 out_unlock:
2969         unlock_extent_cached(io_tree, ordered_extent->file_offset,
2970                              ordered_extent->file_offset +
2971                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
2972 out:
2973         if (root != root->fs_info->tree_root)
2974                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
2975         if (trans)
2976                 btrfs_end_transaction(trans, root);
2977
2978         if (ret || truncated) {
2979                 u64 start, end;
2980
2981                 if (truncated)
2982                         start = ordered_extent->file_offset + logical_len;
2983                 else
2984                         start = ordered_extent->file_offset;
2985                 end = ordered_extent->file_offset + ordered_extent->len - 1;
2986                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
2987
2988                 /* Drop the cache for the part of the extent we didn't write. */
2989                 btrfs_drop_extent_cache(inode, start, end, 0);
2990
2991                 /*
2992                  * If the ordered extent had an IOERR or something else went
2993                  * wrong we need to return the space for this ordered extent
2994                  * back to the allocator.  We only free the extent in the
2995                  * truncated case if we didn't write out the extent at all.
2996                  */
2997                 if ((ret || !logical_len) &&
2998                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2999                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3000                         btrfs_free_reserved_extent(root, ordered_extent->start,
3001                                                    ordered_extent->disk_len, 1);
3002         }
3003
3004
3005         /*
3006          * This needs to be done to make sure anybody waiting knows we are done
3007          * updating everything for this ordered extent.
3008          */
3009         btrfs_remove_ordered_extent(inode, ordered_extent);
3010
3011         /* for snapshot-aware defrag */
3012         if (new) {
3013                 if (ret) {
3014                         free_sa_defrag_extent(new);
3015                         atomic_dec(&root->fs_info->defrag_running);
3016                 } else {
3017                         relink_file_extents(new);
3018                 }
3019         }
3020
3021         /* once for us */
3022         btrfs_put_ordered_extent(ordered_extent);
3023         /* once for the tree */
3024         btrfs_put_ordered_extent(ordered_extent);
3025
3026         return ret;
3027 }
3028
3029 static void finish_ordered_fn(struct btrfs_work *work)
3030 {
3031         struct btrfs_ordered_extent *ordered_extent;
3032         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3033         btrfs_finish_ordered_io(ordered_extent);
3034 }
3035
3036 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3037                                 struct extent_state *state, int uptodate)
3038 {
3039         struct inode *inode = page->mapping->host;
3040         struct btrfs_root *root = BTRFS_I(inode)->root;
3041         struct btrfs_ordered_extent *ordered_extent = NULL;
3042         struct btrfs_workqueue *wq;
3043         btrfs_work_func_t func;
3044
3045         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3046
3047         ClearPagePrivate2(page);
3048         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3049                                             end - start + 1, uptodate))
3050                 return 0;
3051
3052         if (btrfs_is_free_space_inode(inode)) {
3053                 wq = root->fs_info->endio_freespace_worker;
3054                 func = btrfs_freespace_write_helper;
3055         } else {
3056                 wq = root->fs_info->endio_write_workers;
3057                 func = btrfs_endio_write_helper;
3058         }
3059
3060         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3061                         NULL);
3062         btrfs_queue_work(wq, &ordered_extent->work);
3063
3064         return 0;
3065 }
3066
3067 static int __readpage_endio_check(struct inode *inode,
3068                                   struct btrfs_io_bio *io_bio,
3069                                   int icsum, struct page *page,
3070                                   int pgoff, u64 start, size_t len)
3071 {
3072         char *kaddr;
3073         u32 csum_expected;
3074         u32 csum = ~(u32)0;
3075
3076         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3077
3078         kaddr = kmap_atomic(page);
3079         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3080         btrfs_csum_final(csum, (char *)&csum);
3081         if (csum != csum_expected)
3082                 goto zeroit;
3083
3084         kunmap_atomic(kaddr);
3085         return 0;
3086 zeroit:
3087         btrfs_warn_rl(BTRFS_I(inode)->root->fs_info,
3088                 "csum failed ino %llu off %llu csum %u expected csum %u",
3089                            btrfs_ino(inode), start, csum, csum_expected);
3090         memset(kaddr + pgoff, 1, len);
3091         flush_dcache_page(page);
3092         kunmap_atomic(kaddr);
3093         if (csum_expected == 0)
3094                 return 0;
3095         return -EIO;
3096 }
3097
3098 /*
3099  * when reads are done, we need to check csums to verify the data is correct
3100  * if there's a match, we allow the bio to finish.  If not, the code in
3101  * extent_io.c will try to find good copies for us.
3102  */
3103 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3104                                       u64 phy_offset, struct page *page,
3105                                       u64 start, u64 end, int mirror)
3106 {
3107         size_t offset = start - page_offset(page);
3108         struct inode *inode = page->mapping->host;
3109         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3110         struct btrfs_root *root = BTRFS_I(inode)->root;
3111
3112         if (PageChecked(page)) {
3113                 ClearPageChecked(page);
3114                 return 0;
3115         }
3116
3117         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3118                 return 0;
3119
3120         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3121             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3122                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3123                 return 0;
3124         }
3125
3126         phy_offset >>= inode->i_sb->s_blocksize_bits;
3127         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3128                                       start, (size_t)(end - start + 1));
3129 }
3130
3131 void btrfs_add_delayed_iput(struct inode *inode)
3132 {
3133         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3134         struct btrfs_inode *binode = BTRFS_I(inode);
3135
3136         if (atomic_add_unless(&inode->i_count, -1, 1))
3137                 return;
3138
3139         spin_lock(&fs_info->delayed_iput_lock);
3140         if (binode->delayed_iput_count == 0) {
3141                 ASSERT(list_empty(&binode->delayed_iput));
3142                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3143         } else {
3144                 binode->delayed_iput_count++;
3145         }
3146         spin_unlock(&fs_info->delayed_iput_lock);
3147 }
3148
3149 void btrfs_run_delayed_iputs(struct btrfs_root *root)
3150 {
3151         struct btrfs_fs_info *fs_info = root->fs_info;
3152
3153         spin_lock(&fs_info->delayed_iput_lock);
3154         while (!list_empty(&fs_info->delayed_iputs)) {
3155                 struct btrfs_inode *inode;
3156
3157                 inode = list_first_entry(&fs_info->delayed_iputs,
3158                                 struct btrfs_inode, delayed_iput);
3159                 if (inode->delayed_iput_count) {
3160                         inode->delayed_iput_count--;
3161                         list_move_tail(&inode->delayed_iput,
3162                                         &fs_info->delayed_iputs);
3163                 } else {
3164                         list_del_init(&inode->delayed_iput);
3165                 }
3166                 spin_unlock(&fs_info->delayed_iput_lock);
3167                 iput(&inode->vfs_inode);
3168                 spin_lock(&fs_info->delayed_iput_lock);
3169         }
3170         spin_unlock(&fs_info->delayed_iput_lock);
3171 }
3172
3173 /*
3174  * This is called in transaction commit time. If there are no orphan
3175  * files in the subvolume, it removes orphan item and frees block_rsv
3176  * structure.
3177  */
3178 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3179                               struct btrfs_root *root)
3180 {
3181         struct btrfs_block_rsv *block_rsv;
3182         int ret;
3183
3184         if (atomic_read(&root->orphan_inodes) ||
3185             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3186                 return;
3187
3188         spin_lock(&root->orphan_lock);
3189         if (atomic_read(&root->orphan_inodes)) {
3190                 spin_unlock(&root->orphan_lock);
3191                 return;
3192         }
3193
3194         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3195                 spin_unlock(&root->orphan_lock);
3196                 return;
3197         }
3198
3199         block_rsv = root->orphan_block_rsv;
3200         root->orphan_block_rsv = NULL;
3201         spin_unlock(&root->orphan_lock);
3202
3203         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3204             btrfs_root_refs(&root->root_item) > 0) {
3205                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
3206                                             root->root_key.objectid);
3207                 if (ret)
3208                         btrfs_abort_transaction(trans, root, ret);
3209                 else
3210                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3211                                   &root->state);
3212         }
3213
3214         if (block_rsv) {
3215                 WARN_ON(block_rsv->size > 0);
3216                 btrfs_free_block_rsv(root, block_rsv);
3217         }
3218 }
3219
3220 /*
3221  * This creates an orphan entry for the given inode in case something goes
3222  * wrong in the middle of an unlink/truncate.
3223  *
3224  * NOTE: caller of this function should reserve 5 units of metadata for
3225  *       this function.
3226  */
3227 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
3228 {
3229         struct btrfs_root *root = BTRFS_I(inode)->root;
3230         struct btrfs_block_rsv *block_rsv = NULL;
3231         int reserve = 0;
3232         int insert = 0;
3233         int ret;
3234
3235         if (!root->orphan_block_rsv) {
3236                 block_rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
3237                 if (!block_rsv)
3238                         return -ENOMEM;
3239         }
3240
3241         spin_lock(&root->orphan_lock);
3242         if (!root->orphan_block_rsv) {
3243                 root->orphan_block_rsv = block_rsv;
3244         } else if (block_rsv) {
3245                 btrfs_free_block_rsv(root, block_rsv);
3246                 block_rsv = NULL;
3247         }
3248
3249         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3250                               &BTRFS_I(inode)->runtime_flags)) {
3251 #if 0
3252                 /*
3253                  * For proper ENOSPC handling, we should do orphan
3254                  * cleanup when mounting. But this introduces backward
3255                  * compatibility issue.
3256                  */
3257                 if (!xchg(&root->orphan_item_inserted, 1))
3258                         insert = 2;
3259                 else
3260                         insert = 1;
3261 #endif
3262                 insert = 1;
3263                 atomic_inc(&root->orphan_inodes);
3264         }
3265
3266         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3267                               &BTRFS_I(inode)->runtime_flags))
3268                 reserve = 1;
3269         spin_unlock(&root->orphan_lock);
3270
3271         /* grab metadata reservation from transaction handle */
3272         if (reserve) {
3273                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3274                 ASSERT(!ret);
3275                 if (ret) {
3276                         atomic_dec(&root->orphan_inodes);
3277                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3278                                   &BTRFS_I(inode)->runtime_flags);
3279                         if (insert)
3280                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3281                                           &BTRFS_I(inode)->runtime_flags);
3282                         return ret;
3283                 }
3284         }
3285
3286         /* insert an orphan item to track this unlinked/truncated file */
3287         if (insert >= 1) {
3288                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3289                 if (ret) {
3290                         atomic_dec(&root->orphan_inodes);
3291                         if (reserve) {
3292                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3293                                           &BTRFS_I(inode)->runtime_flags);
3294                                 btrfs_orphan_release_metadata(inode);
3295                         }
3296                         if (ret != -EEXIST) {
3297                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3298                                           &BTRFS_I(inode)->runtime_flags);
3299                                 btrfs_abort_transaction(trans, root, ret);
3300                                 return ret;
3301                         }
3302                 }
3303                 ret = 0;
3304         }
3305
3306         /* insert an orphan item to track subvolume contains orphan files */
3307         if (insert >= 2) {
3308                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
3309                                                root->root_key.objectid);
3310                 if (ret && ret != -EEXIST) {
3311                         btrfs_abort_transaction(trans, root, ret);
3312                         return ret;
3313                 }
3314         }
3315         return 0;
3316 }
3317
3318 /*
3319  * We have done the truncate/delete so we can go ahead and remove the orphan
3320  * item for this particular inode.
3321  */
3322 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3323                             struct inode *inode)
3324 {
3325         struct btrfs_root *root = BTRFS_I(inode)->root;
3326         int delete_item = 0;
3327         int release_rsv = 0;
3328         int ret = 0;
3329
3330         spin_lock(&root->orphan_lock);
3331         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3332                                &BTRFS_I(inode)->runtime_flags))
3333                 delete_item = 1;
3334
3335         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3336                                &BTRFS_I(inode)->runtime_flags))
3337                 release_rsv = 1;
3338         spin_unlock(&root->orphan_lock);
3339
3340         if (delete_item) {
3341                 atomic_dec(&root->orphan_inodes);
3342                 if (trans)
3343                         ret = btrfs_del_orphan_item(trans, root,
3344                                                     btrfs_ino(inode));
3345         }
3346
3347         if (release_rsv)
3348                 btrfs_orphan_release_metadata(inode);
3349
3350         return ret;
3351 }
3352
3353 /*
3354  * this cleans up any orphans that may be left on the list from the last use
3355  * of this root.
3356  */
3357 int btrfs_orphan_cleanup(struct btrfs_root *root)
3358 {
3359         struct btrfs_path *path;
3360         struct extent_buffer *leaf;
3361         struct btrfs_key key, found_key;
3362         struct btrfs_trans_handle *trans;
3363         struct inode *inode;
3364         u64 last_objectid = 0;
3365         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3366
3367         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3368                 return 0;
3369
3370         path = btrfs_alloc_path();
3371         if (!path) {
3372                 ret = -ENOMEM;
3373                 goto out;
3374         }
3375         path->reada = READA_BACK;
3376
3377         key.objectid = BTRFS_ORPHAN_OBJECTID;
3378         key.type = BTRFS_ORPHAN_ITEM_KEY;
3379         key.offset = (u64)-1;
3380
3381         while (1) {
3382                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3383                 if (ret < 0)
3384                         goto out;
3385
3386                 /*
3387                  * if ret == 0 means we found what we were searching for, which
3388                  * is weird, but possible, so only screw with path if we didn't
3389                  * find the key and see if we have stuff that matches
3390                  */
3391                 if (ret > 0) {
3392                         ret = 0;
3393                         if (path->slots[0] == 0)
3394                                 break;
3395                         path->slots[0]--;
3396                 }
3397
3398                 /* pull out the item */
3399                 leaf = path->nodes[0];
3400                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3401
3402                 /* make sure the item matches what we want */
3403                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3404                         break;
3405                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3406                         break;
3407
3408                 /* release the path since we're done with it */
3409                 btrfs_release_path(path);
3410
3411                 /*
3412                  * this is where we are basically btrfs_lookup, without the
3413                  * crossing root thing.  we store the inode number in the
3414                  * offset of the orphan item.
3415                  */
3416
3417                 if (found_key.offset == last_objectid) {
3418                         btrfs_err(root->fs_info,
3419                                 "Error removing orphan entry, stopping orphan cleanup");
3420                         ret = -EINVAL;
3421                         goto out;
3422                 }
3423
3424                 last_objectid = found_key.offset;
3425
3426                 found_key.objectid = found_key.offset;
3427                 found_key.type = BTRFS_INODE_ITEM_KEY;
3428                 found_key.offset = 0;
3429                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
3430                 ret = PTR_ERR_OR_ZERO(inode);
3431                 if (ret && ret != -ESTALE)
3432                         goto out;
3433
3434                 if (ret == -ESTALE && root == root->fs_info->tree_root) {
3435                         struct btrfs_root *dead_root;
3436                         struct btrfs_fs_info *fs_info = root->fs_info;
3437                         int is_dead_root = 0;
3438
3439                         /*
3440                          * this is an orphan in the tree root. Currently these
3441                          * could come from 2 sources:
3442                          *  a) a snapshot deletion in progress
3443                          *  b) a free space cache inode
3444                          * We need to distinguish those two, as the snapshot
3445                          * orphan must not get deleted.
3446                          * find_dead_roots already ran before us, so if this
3447                          * is a snapshot deletion, we should find the root
3448                          * in the dead_roots list
3449                          */
3450                         spin_lock(&fs_info->trans_lock);
3451                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3452                                             root_list) {
3453                                 if (dead_root->root_key.objectid ==
3454                                     found_key.objectid) {
3455                                         is_dead_root = 1;
3456                                         break;
3457                                 }
3458                         }
3459                         spin_unlock(&fs_info->trans_lock);
3460                         if (is_dead_root) {
3461                                 /* prevent this orphan from being found again */
3462                                 key.offset = found_key.objectid - 1;
3463                                 continue;
3464                         }
3465                 }
3466                 /*
3467                  * Inode is already gone but the orphan item is still there,
3468                  * kill the orphan item.
3469                  */
3470                 if (ret == -ESTALE) {
3471                         trans = btrfs_start_transaction(root, 1);
3472                         if (IS_ERR(trans)) {
3473                                 ret = PTR_ERR(trans);
3474                                 goto out;
3475                         }
3476                         btrfs_debug(root->fs_info, "auto deleting %Lu",
3477                                 found_key.objectid);
3478                         ret = btrfs_del_orphan_item(trans, root,
3479                                                     found_key.objectid);
3480                         btrfs_end_transaction(trans, root);
3481                         if (ret)
3482                                 goto out;
3483                         continue;
3484                 }
3485
3486                 /*
3487                  * add this inode to the orphan list so btrfs_orphan_del does
3488                  * the proper thing when we hit it
3489                  */
3490                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3491                         &BTRFS_I(inode)->runtime_flags);
3492                 atomic_inc(&root->orphan_inodes);
3493
3494                 /* if we have links, this was a truncate, lets do that */
3495                 if (inode->i_nlink) {
3496                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3497                                 iput(inode);
3498                                 continue;
3499                         }
3500                         nr_truncate++;
3501
3502                         /* 1 for the orphan item deletion. */
3503                         trans = btrfs_start_transaction(root, 1);
3504                         if (IS_ERR(trans)) {
3505                                 iput(inode);
3506                                 ret = PTR_ERR(trans);
3507                                 goto out;
3508                         }
3509                         ret = btrfs_orphan_add(trans, inode);
3510                         btrfs_end_transaction(trans, root);
3511                         if (ret) {
3512                                 iput(inode);
3513                                 goto out;
3514                         }
3515
3516                         ret = btrfs_truncate(inode);
3517                         if (ret)
3518                                 btrfs_orphan_del(NULL, inode);
3519                 } else {
3520                         nr_unlink++;
3521                 }
3522
3523                 /* this will do delete_inode and everything for us */
3524                 iput(inode);
3525                 if (ret)
3526                         goto out;
3527         }
3528         /* release the path since we're done with it */
3529         btrfs_release_path(path);
3530
3531         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3532
3533         if (root->orphan_block_rsv)
3534                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
3535                                         (u64)-1);
3536
3537         if (root->orphan_block_rsv ||
3538             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3539                 trans = btrfs_join_transaction(root);
3540                 if (!IS_ERR(trans))
3541                         btrfs_end_transaction(trans, root);
3542         }
3543
3544         if (nr_unlink)
3545                 btrfs_debug(root->fs_info, "unlinked %d orphans", nr_unlink);
3546         if (nr_truncate)
3547                 btrfs_debug(root->fs_info, "truncated %d orphans", nr_truncate);
3548
3549 out:
3550         if (ret)
3551                 btrfs_err(root->fs_info,
3552                         "could not do orphan cleanup %d", ret);
3553         btrfs_free_path(path);
3554         return ret;
3555 }
3556
3557 /*
3558  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3559  * don't find any xattrs, we know there can't be any acls.
3560  *
3561  * slot is the slot the inode is in, objectid is the objectid of the inode
3562  */
3563 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3564                                           int slot, u64 objectid,
3565                                           int *first_xattr_slot)
3566 {
3567         u32 nritems = btrfs_header_nritems(leaf);
3568         struct btrfs_key found_key;
3569         static u64 xattr_access = 0;
3570         static u64 xattr_default = 0;
3571         int scanned = 0;
3572
3573         if (!xattr_access) {
3574                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3575                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3576                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3577                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3578         }
3579
3580         slot++;
3581         *first_xattr_slot = -1;
3582         while (slot < nritems) {
3583                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3584
3585                 /* we found a different objectid, there must not be acls */
3586                 if (found_key.objectid != objectid)
3587                         return 0;
3588
3589                 /* we found an xattr, assume we've got an acl */
3590                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3591                         if (*first_xattr_slot == -1)
3592                                 *first_xattr_slot = slot;
3593                         if (found_key.offset == xattr_access ||
3594                             found_key.offset == xattr_default)
3595                                 return 1;
3596                 }
3597
3598                 /*
3599                  * we found a key greater than an xattr key, there can't
3600                  * be any acls later on
3601                  */
3602                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3603                         return 0;
3604
3605                 slot++;
3606                 scanned++;
3607
3608                 /*
3609                  * it goes inode, inode backrefs, xattrs, extents,
3610                  * so if there are a ton of hard links to an inode there can
3611                  * be a lot of backrefs.  Don't waste time searching too hard,
3612                  * this is just an optimization
3613                  */
3614                 if (scanned >= 8)
3615                         break;
3616         }
3617         /* we hit the end of the leaf before we found an xattr or
3618          * something larger than an xattr.  We have to assume the inode
3619          * has acls
3620          */
3621         if (*first_xattr_slot == -1)
3622                 *first_xattr_slot = slot;
3623         return 1;
3624 }
3625
3626 /*
3627  * read an inode from the btree into the in-memory inode
3628  */
3629 static void btrfs_read_locked_inode(struct inode *inode)
3630 {
3631         struct btrfs_path *path;
3632         struct extent_buffer *leaf;
3633         struct btrfs_inode_item *inode_item;
3634         struct btrfs_root *root = BTRFS_I(inode)->root;
3635         struct btrfs_key location;
3636         unsigned long ptr;
3637         int maybe_acls;
3638         u32 rdev;
3639         int ret;
3640         bool filled = false;
3641         int first_xattr_slot;
3642
3643         ret = btrfs_fill_inode(inode, &rdev);
3644         if (!ret)
3645                 filled = true;
3646
3647         path = btrfs_alloc_path();
3648         if (!path)
3649                 goto make_bad;
3650
3651         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3652
3653         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3654         if (ret)
3655                 goto make_bad;
3656
3657         leaf = path->nodes[0];
3658
3659         if (filled)
3660                 goto cache_index;
3661
3662         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3663                                     struct btrfs_inode_item);
3664         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3665         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3666         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3667         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3668         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
3669
3670         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3671         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3672
3673         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3674         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3675
3676         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3677         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3678
3679         BTRFS_I(inode)->i_otime.tv_sec =
3680                 btrfs_timespec_sec(leaf, &inode_item->otime);
3681         BTRFS_I(inode)->i_otime.tv_nsec =
3682                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3683
3684         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3685         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3686         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3687
3688         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3689         inode->i_generation = BTRFS_I(inode)->generation;
3690         inode->i_rdev = 0;
3691         rdev = btrfs_inode_rdev(leaf, inode_item);
3692
3693         BTRFS_I(inode)->index_cnt = (u64)-1;
3694         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3695
3696 cache_index:
3697         /*
3698          * If we were modified in the current generation and evicted from memory
3699          * and then re-read we need to do a full sync since we don't have any
3700          * idea about which extents were modified before we were evicted from
3701          * cache.
3702          *
3703          * This is required for both inode re-read from disk and delayed inode
3704          * in delayed_nodes_tree.
3705          */
3706         if (BTRFS_I(inode)->last_trans == root->fs_info->generation)
3707                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3708                         &BTRFS_I(inode)->runtime_flags);
3709
3710         /*
3711          * We don't persist the id of the transaction where an unlink operation
3712          * against the inode was last made. So here we assume the inode might
3713          * have been evicted, and therefore the exact value of last_unlink_trans
3714          * lost, and set it to last_trans to avoid metadata inconsistencies
3715          * between the inode and its parent if the inode is fsync'ed and the log
3716          * replayed. For example, in the scenario:
3717          *
3718          * touch mydir/foo
3719          * ln mydir/foo mydir/bar
3720          * sync
3721          * unlink mydir/bar
3722          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3723          * xfs_io -c fsync mydir/foo
3724          * <power failure>
3725          * mount fs, triggers fsync log replay
3726          *
3727          * We must make sure that when we fsync our inode foo we also log its
3728          * parent inode, otherwise after log replay the parent still has the
3729          * dentry with the "bar" name but our inode foo has a link count of 1
3730          * and doesn't have an inode ref with the name "bar" anymore.
3731          *
3732          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3733          * but it guarantees correctness at the expense of occasional full
3734          * transaction commits on fsync if our inode is a directory, or if our
3735          * inode is not a directory, logging its parent unnecessarily.
3736          */
3737         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3738
3739         path->slots[0]++;
3740         if (inode->i_nlink != 1 ||
3741             path->slots[0] >= btrfs_header_nritems(leaf))
3742                 goto cache_acl;
3743
3744         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3745         if (location.objectid != btrfs_ino(inode))
3746                 goto cache_acl;
3747
3748         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3749         if (location.type == BTRFS_INODE_REF_KEY) {
3750                 struct btrfs_inode_ref *ref;
3751
3752                 ref = (struct btrfs_inode_ref *)ptr;
3753                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3754         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3755                 struct btrfs_inode_extref *extref;
3756
3757                 extref = (struct btrfs_inode_extref *)ptr;
3758                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3759                                                                      extref);
3760         }
3761 cache_acl:
3762         /*
3763          * try to precache a NULL acl entry for files that don't have
3764          * any xattrs or acls
3765          */
3766         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3767                                            btrfs_ino(inode), &first_xattr_slot);
3768         if (first_xattr_slot != -1) {
3769                 path->slots[0] = first_xattr_slot;
3770                 ret = btrfs_load_inode_props(inode, path);
3771                 if (ret)
3772                         btrfs_err(root->fs_info,
3773                                   "error loading props for ino %llu (root %llu): %d",
3774                                   btrfs_ino(inode),
3775                                   root->root_key.objectid, ret);
3776         }
3777         btrfs_free_path(path);
3778
3779         if (!maybe_acls)
3780                 cache_no_acl(inode);
3781
3782         switch (inode->i_mode & S_IFMT) {
3783         case S_IFREG:
3784                 inode->i_mapping->a_ops = &btrfs_aops;
3785                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3786                 inode->i_fop = &btrfs_file_operations;
3787                 inode->i_op = &btrfs_file_inode_operations;
3788                 break;
3789         case S_IFDIR:
3790                 inode->i_fop = &btrfs_dir_file_operations;
3791                 if (root == root->fs_info->tree_root)
3792                         inode->i_op = &btrfs_dir_ro_inode_operations;
3793                 else
3794                         inode->i_op = &btrfs_dir_inode_operations;
3795                 break;
3796         case S_IFLNK:
3797                 inode->i_op = &btrfs_symlink_inode_operations;
3798                 inode_nohighmem(inode);
3799                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3800                 break;
3801         default:
3802                 inode->i_op = &btrfs_special_inode_operations;
3803                 init_special_inode(inode, inode->i_mode, rdev);
3804                 break;
3805         }
3806
3807         btrfs_update_iflags(inode);
3808         return;
3809
3810 make_bad:
3811         btrfs_free_path(path);
3812         make_bad_inode(inode);
3813 }
3814
3815 /*
3816  * given a leaf and an inode, copy the inode fields into the leaf
3817  */
3818 static void fill_inode_item(struct btrfs_trans_handle *trans,
3819                             struct extent_buffer *leaf,
3820                             struct btrfs_inode_item *item,
3821                             struct inode *inode)
3822 {
3823         struct btrfs_map_token token;
3824
3825         btrfs_init_map_token(&token);
3826
3827         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3828         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3829         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3830                                    &token);
3831         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3832         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3833
3834         btrfs_set_token_timespec_sec(leaf, &item->atime,
3835                                      inode->i_atime.tv_sec, &token);
3836         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3837                                       inode->i_atime.tv_nsec, &token);
3838
3839         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3840                                      inode->i_mtime.tv_sec, &token);
3841         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3842                                       inode->i_mtime.tv_nsec, &token);
3843
3844         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3845                                      inode->i_ctime.tv_sec, &token);
3846         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3847                                       inode->i_ctime.tv_nsec, &token);
3848
3849         btrfs_set_token_timespec_sec(leaf, &item->otime,
3850                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3851         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3852                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3853
3854         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3855                                      &token);
3856         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3857                                          &token);
3858         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3859         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3860         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3861         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3862         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3863 }
3864
3865 /*
3866  * copy everything in the in-memory inode into the btree.
3867  */
3868 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3869                                 struct btrfs_root *root, struct inode *inode)
3870 {
3871         struct btrfs_inode_item *inode_item;
3872         struct btrfs_path *path;
3873         struct extent_buffer *leaf;
3874         int ret;
3875
3876         path = btrfs_alloc_path();
3877         if (!path)
3878                 return -ENOMEM;
3879
3880         path->leave_spinning = 1;
3881         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3882                                  1);
3883         if (ret) {
3884                 if (ret > 0)
3885                         ret = -ENOENT;
3886                 goto failed;
3887         }
3888
3889         leaf = path->nodes[0];
3890         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3891                                     struct btrfs_inode_item);
3892
3893         fill_inode_item(trans, leaf, inode_item, inode);
3894         btrfs_mark_buffer_dirty(leaf);
3895         btrfs_set_inode_last_trans(trans, inode);
3896         ret = 0;
3897 failed:
3898         btrfs_free_path(path);
3899         return ret;
3900 }
3901
3902 /*
3903  * copy everything in the in-memory inode into the btree.
3904  */
3905 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3906                                 struct btrfs_root *root, struct inode *inode)
3907 {
3908         int ret;
3909
3910         /*
3911          * If the inode is a free space inode, we can deadlock during commit
3912          * if we put it into the delayed code.
3913          *
3914          * The data relocation inode should also be directly updated
3915          * without delay
3916          */
3917         if (!btrfs_is_free_space_inode(inode)
3918             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3919             && !root->fs_info->log_root_recovering) {
3920                 btrfs_update_root_times(trans, root);
3921
3922                 ret = btrfs_delayed_update_inode(trans, root, inode);
3923                 if (!ret)
3924                         btrfs_set_inode_last_trans(trans, inode);
3925                 return ret;
3926         }
3927
3928         return btrfs_update_inode_item(trans, root, inode);
3929 }
3930
3931 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3932                                          struct btrfs_root *root,
3933                                          struct inode *inode)
3934 {
3935         int ret;
3936
3937         ret = btrfs_update_inode(trans, root, inode);
3938         if (ret == -ENOSPC)
3939                 return btrfs_update_inode_item(trans, root, inode);
3940         return ret;
3941 }
3942
3943 /*
3944  * unlink helper that gets used here in inode.c and in the tree logging
3945  * recovery code.  It remove a link in a directory with a given name, and
3946  * also drops the back refs in the inode to the directory
3947  */
3948 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3949                                 struct btrfs_root *root,
3950                                 struct inode *dir, struct inode *inode,
3951                                 const char *name, int name_len)
3952 {
3953         struct btrfs_path *path;
3954         int ret = 0;
3955         struct extent_buffer *leaf;
3956         struct btrfs_dir_item *di;
3957         struct btrfs_key key;
3958         u64 index;
3959         u64 ino = btrfs_ino(inode);
3960         u64 dir_ino = btrfs_ino(dir);
3961
3962         path = btrfs_alloc_path();
3963         if (!path) {
3964                 ret = -ENOMEM;
3965                 goto out;
3966         }
3967
3968         path->leave_spinning = 1;
3969         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3970                                     name, name_len, -1);
3971         if (IS_ERR(di)) {
3972                 ret = PTR_ERR(di);
3973                 goto err;
3974         }
3975         if (!di) {
3976                 ret = -ENOENT;
3977                 goto err;
3978         }
3979         leaf = path->nodes[0];
3980         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3981         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3982         if (ret)
3983                 goto err;
3984         btrfs_release_path(path);
3985
3986         /*
3987          * If we don't have dir index, we have to get it by looking up
3988          * the inode ref, since we get the inode ref, remove it directly,
3989          * it is unnecessary to do delayed deletion.
3990          *
3991          * But if we have dir index, needn't search inode ref to get it.
3992          * Since the inode ref is close to the inode item, it is better
3993          * that we delay to delete it, and just do this deletion when
3994          * we update the inode item.
3995          */
3996         if (BTRFS_I(inode)->dir_index) {
3997                 ret = btrfs_delayed_delete_inode_ref(inode);
3998                 if (!ret) {
3999                         index = BTRFS_I(inode)->dir_index;
4000                         goto skip_backref;
4001                 }
4002         }
4003
4004         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4005                                   dir_ino, &index);
4006         if (ret) {
4007                 btrfs_info(root->fs_info,
4008                         "failed to delete reference to %.*s, inode %llu parent %llu",
4009                         name_len, name, ino, dir_ino);
4010                 btrfs_abort_transaction(trans, root, ret);
4011                 goto err;
4012         }
4013 skip_backref:
4014         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4015         if (ret) {
4016                 btrfs_abort_transaction(trans, root, ret);
4017                 goto err;
4018         }
4019
4020         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
4021                                          inode, dir_ino);
4022         if (ret != 0 && ret != -ENOENT) {
4023                 btrfs_abort_transaction(trans, root, ret);
4024                 goto err;
4025         }
4026
4027         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
4028                                            dir, index);
4029         if (ret == -ENOENT)
4030                 ret = 0;
4031         else if (ret)
4032                 btrfs_abort_transaction(trans, root, ret);
4033 err:
4034         btrfs_free_path(path);
4035         if (ret)
4036                 goto out;
4037
4038         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4039         inode_inc_iversion(inode);
4040         inode_inc_iversion(dir);
4041         inode->i_ctime = dir->i_mtime =
4042                 dir->i_ctime = current_fs_time(inode->i_sb);
4043         ret = btrfs_update_inode(trans, root, dir);
4044 out:
4045         return ret;
4046 }
4047
4048 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4049                        struct btrfs_root *root,
4050                        struct inode *dir, struct inode *inode,
4051                        const char *name, int name_len)
4052 {
4053         int ret;
4054         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4055         if (!ret) {
4056                 drop_nlink(inode);
4057                 ret = btrfs_update_inode(trans, root, inode);
4058         }
4059         return ret;
4060 }
4061
4062 /*
4063  * helper to start transaction for unlink and rmdir.
4064  *
4065  * unlink and rmdir are special in btrfs, they do not always free space, so
4066  * if we cannot make our reservations the normal way try and see if there is
4067  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4068  * allow the unlink to occur.
4069  */
4070 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4071 {
4072         struct btrfs_root *root = BTRFS_I(dir)->root;
4073
4074         /*
4075          * 1 for the possible orphan item
4076          * 1 for the dir item
4077          * 1 for the dir index
4078          * 1 for the inode ref
4079          * 1 for the inode
4080          */
4081         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4082 }
4083
4084 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4085 {
4086         struct btrfs_root *root = BTRFS_I(dir)->root;
4087         struct btrfs_trans_handle *trans;
4088         struct inode *inode = d_inode(dentry);
4089         int ret;
4090
4091         trans = __unlink_start_trans(dir);
4092         if (IS_ERR(trans))
4093                 return PTR_ERR(trans);
4094
4095         btrfs_record_unlink_dir(trans, dir, d_inode(dentry), 0);
4096
4097         ret = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4098                                  dentry->d_name.name, dentry->d_name.len);
4099         if (ret)
4100                 goto out;
4101
4102         if (inode->i_nlink == 0) {
4103                 ret = btrfs_orphan_add(trans, inode);
4104                 if (ret)
4105                         goto out;
4106         }
4107
4108 out:
4109         btrfs_end_transaction(trans, root);
4110         btrfs_btree_balance_dirty(root);
4111         return ret;
4112 }
4113
4114 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4115                         struct btrfs_root *root,
4116                         struct inode *dir, u64 objectid,
4117                         const char *name, int name_len)
4118 {
4119         struct btrfs_path *path;
4120         struct extent_buffer *leaf;
4121         struct btrfs_dir_item *di;
4122         struct btrfs_key key;
4123         u64 index;
4124         int ret;
4125         u64 dir_ino = btrfs_ino(dir);
4126
4127         path = btrfs_alloc_path();
4128         if (!path)
4129                 return -ENOMEM;
4130
4131         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4132                                    name, name_len, -1);
4133         if (IS_ERR_OR_NULL(di)) {
4134                 if (!di)
4135                         ret = -ENOENT;
4136                 else
4137                         ret = PTR_ERR(di);
4138                 goto out;
4139         }
4140
4141         leaf = path->nodes[0];
4142         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4143         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4144         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4145         if (ret) {
4146                 btrfs_abort_transaction(trans, root, ret);
4147                 goto out;
4148         }
4149         btrfs_release_path(path);
4150
4151         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
4152                                  objectid, root->root_key.objectid,
4153                                  dir_ino, &index, name, name_len);
4154         if (ret < 0) {
4155                 if (ret != -ENOENT) {
4156                         btrfs_abort_transaction(trans, root, ret);
4157                         goto out;
4158                 }
4159                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4160                                                  name, name_len);
4161                 if (IS_ERR_OR_NULL(di)) {
4162                         if (!di)
4163                                 ret = -ENOENT;
4164                         else
4165                                 ret = PTR_ERR(di);
4166                         btrfs_abort_transaction(trans, root, ret);
4167                         goto out;
4168                 }
4169
4170                 leaf = path->nodes[0];
4171                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4172                 btrfs_release_path(path);
4173                 index = key.offset;
4174         }
4175         btrfs_release_path(path);
4176
4177         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4178         if (ret) {
4179                 btrfs_abort_transaction(trans, root, ret);
4180                 goto out;
4181         }
4182
4183         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4184         inode_inc_iversion(dir);
4185         dir->i_mtime = dir->i_ctime = current_fs_time(dir->i_sb);
4186         ret = btrfs_update_inode_fallback(trans, root, dir);
4187         if (ret)
4188                 btrfs_abort_transaction(trans, root, ret);
4189 out:
4190         btrfs_free_path(path);
4191         return ret;
4192 }
4193
4194 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4195 {
4196         struct inode *inode = d_inode(dentry);
4197         int err = 0;
4198         struct btrfs_root *root = BTRFS_I(dir)->root;
4199         struct btrfs_trans_handle *trans;
4200
4201         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4202                 return -ENOTEMPTY;
4203         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
4204                 return -EPERM;
4205
4206         trans = __unlink_start_trans(dir);
4207         if (IS_ERR(trans))
4208                 return PTR_ERR(trans);
4209
4210         if (unlikely(btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4211                 err = btrfs_unlink_subvol(trans, root, dir,
4212                                           BTRFS_I(inode)->location.objectid,
4213                                           dentry->d_name.name,
4214                                           dentry->d_name.len);
4215                 goto out;
4216         }
4217
4218         err = btrfs_orphan_add(trans, inode);
4219         if (err)
4220                 goto out;
4221
4222         /* now the directory is empty */
4223         err = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4224                                  dentry->d_name.name, dentry->d_name.len);
4225         if (!err)
4226                 btrfs_i_size_write(inode, 0);
4227 out:
4228         btrfs_end_transaction(trans, root);
4229         btrfs_btree_balance_dirty(root);
4230
4231         return err;
4232 }
4233
4234 static int truncate_space_check(struct btrfs_trans_handle *trans,
4235                                 struct btrfs_root *root,
4236                                 u64 bytes_deleted)
4237 {
4238         int ret;
4239
4240         /*
4241          * This is only used to apply pressure to the enospc system, we don't
4242          * intend to use this reservation at all.
4243          */
4244         bytes_deleted = btrfs_csum_bytes_to_leaves(root, bytes_deleted);
4245         bytes_deleted *= root->nodesize;
4246         ret = btrfs_block_rsv_add(root, &root->fs_info->trans_block_rsv,
4247                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4248         if (!ret) {
4249                 trace_btrfs_space_reservation(root->fs_info, "transaction",
4250                                               trans->transid,
4251                                               bytes_deleted, 1);
4252                 trans->bytes_reserved += bytes_deleted;
4253         }
4254         return ret;
4255
4256 }
4257
4258 static int truncate_inline_extent(struct inode *inode,
4259                                   struct btrfs_path *path,
4260                                   struct btrfs_key *found_key,
4261                                   const u64 item_end,
4262                                   const u64 new_size)
4263 {
4264         struct extent_buffer *leaf = path->nodes[0];
4265         int slot = path->slots[0];
4266         struct btrfs_file_extent_item *fi;
4267         u32 size = (u32)(new_size - found_key->offset);
4268         struct btrfs_root *root = BTRFS_I(inode)->root;
4269
4270         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4271
4272         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4273                 loff_t offset = new_size;
4274                 loff_t page_end = ALIGN(offset, PAGE_SIZE);
4275
4276                 /*
4277                  * Zero out the remaining of the last page of our inline extent,
4278                  * instead of directly truncating our inline extent here - that
4279                  * would be much more complex (decompressing all the data, then
4280                  * compressing the truncated data, which might be bigger than
4281                  * the size of the inline extent, resize the extent, etc).
4282                  * We release the path because to get the page we might need to
4283                  * read the extent item from disk (data not in the page cache).
4284                  */
4285                 btrfs_release_path(path);
4286                 return btrfs_truncate_block(inode, offset, page_end - offset,
4287                                         0);
4288         }
4289
4290         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4291         size = btrfs_file_extent_calc_inline_size(size);
4292         btrfs_truncate_item(root, path, size, 1);
4293
4294         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4295                 inode_sub_bytes(inode, item_end + 1 - new_size);
4296
4297         return 0;
4298 }
4299
4300 /*
4301  * this can truncate away extent items, csum items and directory items.
4302  * It starts at a high offset and removes keys until it can't find
4303  * any higher than new_size
4304  *
4305  * csum items that cross the new i_size are truncated to the new size
4306  * as well.
4307  *
4308  * min_type is the minimum key type to truncate down to.  If set to 0, this
4309  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4310  */
4311 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4312                                struct btrfs_root *root,
4313                                struct inode *inode,
4314                                u64 new_size, u32 min_type)
4315 {
4316         struct btrfs_path *path;
4317         struct extent_buffer *leaf;
4318         struct btrfs_file_extent_item *fi;
4319         struct btrfs_key key;
4320         struct btrfs_key found_key;
4321         u64 extent_start = 0;
4322         u64 extent_num_bytes = 0;
4323         u64 extent_offset = 0;
4324         u64 item_end = 0;
4325         u64 last_size = new_size;
4326         u32 found_type = (u8)-1;
4327         int found_extent;
4328         int del_item;
4329         int pending_del_nr = 0;
4330         int pending_del_slot = 0;
4331         int extent_type = -1;
4332         int ret;
4333         int err = 0;
4334         u64 ino = btrfs_ino(inode);
4335         u64 bytes_deleted = 0;
4336         bool be_nice = 0;
4337         bool should_throttle = 0;
4338         bool should_end = 0;
4339
4340         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4341
4342         /*
4343          * for non-free space inodes and ref cows, we want to back off from
4344          * time to time
4345          */
4346         if (!btrfs_is_free_space_inode(inode) &&
4347             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4348                 be_nice = 1;
4349
4350         path = btrfs_alloc_path();
4351         if (!path)
4352                 return -ENOMEM;
4353         path->reada = READA_BACK;
4354
4355         /*
4356          * We want to drop from the next block forward in case this new size is
4357          * not block aligned since we will be keeping the last block of the
4358          * extent just the way it is.
4359          */
4360         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4361             root == root->fs_info->tree_root)
4362                 btrfs_drop_extent_cache(inode, ALIGN(new_size,
4363                                         root->sectorsize), (u64)-1, 0);
4364
4365         /*
4366          * This function is also used to drop the items in the log tree before
4367          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4368          * it is used to drop the loged items. So we shouldn't kill the delayed
4369          * items.
4370          */
4371         if (min_type == 0 && root == BTRFS_I(inode)->root)
4372                 btrfs_kill_delayed_inode_items(inode);
4373
4374         key.objectid = ino;
4375         key.offset = (u64)-1;
4376         key.type = (u8)-1;
4377
4378 search_again:
4379         /*
4380          * with a 16K leaf size and 128MB extents, you can actually queue
4381          * up a huge file in a single leaf.  Most of the time that
4382          * bytes_deleted is > 0, it will be huge by the time we get here
4383          */
4384         if (be_nice && bytes_deleted > SZ_32M) {
4385                 if (btrfs_should_end_transaction(trans, root)) {
4386                         err = -EAGAIN;
4387                         goto error;
4388                 }
4389         }
4390
4391
4392         path->leave_spinning = 1;
4393         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4394         if (ret < 0) {
4395                 err = ret;
4396                 goto out;
4397         }
4398
4399         if (ret > 0) {
4400                 /* there are no items in the tree for us to truncate, we're
4401                  * done
4402                  */
4403                 if (path->slots[0] == 0)
4404                         goto out;
4405                 path->slots[0]--;
4406         }
4407
4408         while (1) {
4409                 fi = NULL;
4410                 leaf = path->nodes[0];
4411                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4412                 found_type = found_key.type;
4413
4414                 if (found_key.objectid != ino)
4415                         break;
4416
4417                 if (found_type < min_type)
4418                         break;
4419
4420                 item_end = found_key.offset;
4421                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4422                         fi = btrfs_item_ptr(leaf, path->slots[0],
4423                                             struct btrfs_file_extent_item);
4424                         extent_type = btrfs_file_extent_type(leaf, fi);
4425                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4426                                 item_end +=
4427                                     btrfs_file_extent_num_bytes(leaf, fi);
4428                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4429                                 item_end += btrfs_file_extent_inline_len(leaf,
4430                                                          path->slots[0], fi);
4431                         }
4432                         item_end--;
4433                 }
4434                 if (found_type > min_type) {
4435                         del_item = 1;
4436                 } else {
4437                         if (item_end < new_size)
4438                                 break;
4439                         if (found_key.offset >= new_size)
4440                                 del_item = 1;
4441                         else
4442                                 del_item = 0;
4443                 }
4444                 found_extent = 0;
4445                 /* FIXME, shrink the extent if the ref count is only 1 */
4446                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4447                         goto delete;
4448
4449                 if (del_item)
4450                         last_size = found_key.offset;
4451                 else
4452                         last_size = new_size;
4453
4454                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4455                         u64 num_dec;
4456                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4457                         if (!del_item) {
4458                                 u64 orig_num_bytes =
4459                                         btrfs_file_extent_num_bytes(leaf, fi);
4460                                 extent_num_bytes = ALIGN(new_size -
4461                                                 found_key.offset,
4462                                                 root->sectorsize);
4463                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4464                                                          extent_num_bytes);
4465                                 num_dec = (orig_num_bytes -
4466                                            extent_num_bytes);
4467                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4468                                              &root->state) &&
4469                                     extent_start != 0)
4470                                         inode_sub_bytes(inode, num_dec);
4471                                 btrfs_mark_buffer_dirty(leaf);
4472                         } else {
4473                                 extent_num_bytes =
4474                                         btrfs_file_extent_disk_num_bytes(leaf,
4475                                                                          fi);
4476                                 extent_offset = found_key.offset -
4477                                         btrfs_file_extent_offset(leaf, fi);
4478
4479                                 /* FIXME blocksize != 4096 */
4480                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4481                                 if (extent_start != 0) {
4482                                         found_extent = 1;
4483                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4484                                                      &root->state))
4485                                                 inode_sub_bytes(inode, num_dec);
4486                                 }
4487                         }
4488                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4489                         /*
4490                          * we can't truncate inline items that have had
4491                          * special encodings
4492                          */
4493                         if (!del_item &&
4494                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4495                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4496
4497                                 /*
4498                                  * Need to release path in order to truncate a
4499                                  * compressed extent. So delete any accumulated
4500                                  * extent items so far.
4501                                  */
4502                                 if (btrfs_file_extent_compression(leaf, fi) !=
4503                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4504                                         err = btrfs_del_items(trans, root, path,
4505                                                               pending_del_slot,
4506                                                               pending_del_nr);
4507                                         if (err) {
4508                                                 btrfs_abort_transaction(trans,
4509                                                                         root,
4510                                                                         err);
4511                                                 goto error;
4512                                         }
4513                                         pending_del_nr = 0;
4514                                 }
4515
4516                                 err = truncate_inline_extent(inode, path,
4517                                                              &found_key,
4518                                                              item_end,
4519                                                              new_size);
4520                                 if (err) {
4521                                         btrfs_abort_transaction(trans,
4522                                                                 root, err);
4523                                         goto error;
4524                                 }
4525                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4526                                             &root->state)) {
4527                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4528                         }
4529                 }
4530 delete:
4531                 if (del_item) {
4532                         if (!pending_del_nr) {
4533                                 /* no pending yet, add ourselves */
4534                                 pending_del_slot = path->slots[0];
4535                                 pending_del_nr = 1;
4536                         } else if (pending_del_nr &&
4537                                    path->slots[0] + 1 == pending_del_slot) {
4538                                 /* hop on the pending chunk */
4539                                 pending_del_nr++;
4540                                 pending_del_slot = path->slots[0];
4541                         } else {
4542                                 BUG();
4543                         }
4544                 } else {
4545                         break;
4546                 }
4547                 should_throttle = 0;
4548
4549                 if (found_extent &&
4550                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4551                      root == root->fs_info->tree_root)) {
4552                         btrfs_set_path_blocking(path);
4553                         bytes_deleted += extent_num_bytes;
4554                         ret = btrfs_free_extent(trans, root, extent_start,
4555                                                 extent_num_bytes, 0,
4556                                                 btrfs_header_owner(leaf),
4557                                                 ino, extent_offset);
4558                         BUG_ON(ret);
4559                         if (btrfs_should_throttle_delayed_refs(trans, root))
4560                                 btrfs_async_run_delayed_refs(root,
4561                                                              trans->transid,
4562                                         trans->delayed_ref_updates * 2, 0);
4563                         if (be_nice) {
4564                                 if (truncate_space_check(trans, root,
4565                                                          extent_num_bytes)) {
4566                                         should_end = 1;
4567                                 }
4568                                 if (btrfs_should_throttle_delayed_refs(trans,
4569                                                                        root)) {
4570                                         should_throttle = 1;
4571                                 }
4572                         }
4573                 }
4574
4575                 if (found_type == BTRFS_INODE_ITEM_KEY)
4576                         break;
4577
4578                 if (path->slots[0] == 0 ||
4579                     path->slots[0] != pending_del_slot ||
4580                     should_throttle || should_end) {
4581                         if (pending_del_nr) {
4582                                 ret = btrfs_del_items(trans, root, path,
4583                                                 pending_del_slot,
4584                                                 pending_del_nr);
4585                                 if (ret) {
4586                                         btrfs_abort_transaction(trans,
4587                                                                 root, ret);
4588                                         goto error;
4589                                 }
4590                                 pending_del_nr = 0;
4591                         }
4592                         btrfs_release_path(path);
4593                         if (should_throttle) {
4594                                 unsigned long updates = trans->delayed_ref_updates;
4595                                 if (updates) {
4596                                         trans->delayed_ref_updates = 0;
4597                                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4598                                         if (ret && !err)
4599                                                 err = ret;
4600                                 }
4601                         }
4602                         /*
4603                          * if we failed to refill our space rsv, bail out
4604                          * and let the transaction restart
4605                          */
4606                         if (should_end) {
4607                                 err = -EAGAIN;
4608                                 goto error;
4609                         }
4610                         goto search_again;
4611                 } else {
4612                         path->slots[0]--;
4613                 }
4614         }
4615 out:
4616         if (pending_del_nr) {
4617                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4618                                       pending_del_nr);
4619                 if (ret)
4620                         btrfs_abort_transaction(trans, root, ret);
4621         }
4622 error:
4623         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4624                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4625
4626         btrfs_free_path(path);
4627
4628         if (be_nice && bytes_deleted > SZ_32M) {
4629                 unsigned long updates = trans->delayed_ref_updates;
4630                 if (updates) {
4631                         trans->delayed_ref_updates = 0;
4632                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4633                         if (ret && !err)
4634                                 err = ret;
4635                 }
4636         }
4637         return err;
4638 }
4639
4640 /*
4641  * btrfs_truncate_block - read, zero a chunk and write a block
4642  * @inode - inode that we're zeroing
4643  * @from - the offset to start zeroing
4644  * @len - the length to zero, 0 to zero the entire range respective to the
4645  *      offset
4646  * @front - zero up to the offset instead of from the offset on
4647  *
4648  * This will find the block for the "from" offset and cow the block and zero the
4649  * part we want to zero.  This is used with truncate and hole punching.
4650  */
4651 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4652                         int front)
4653 {
4654         struct address_space *mapping = inode->i_mapping;
4655         struct btrfs_root *root = BTRFS_I(inode)->root;
4656         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4657         struct btrfs_ordered_extent *ordered;
4658         struct extent_state *cached_state = NULL;
4659         char *kaddr;
4660         u32 blocksize = root->sectorsize;
4661         pgoff_t index = from >> PAGE_SHIFT;
4662         unsigned offset = from & (blocksize - 1);
4663         struct page *page;
4664         gfp_t mask = btrfs_alloc_write_mask(mapping);
4665         int ret = 0;
4666         u64 block_start;
4667         u64 block_end;
4668
4669         if ((offset & (blocksize - 1)) == 0 &&
4670             (!len || ((len & (blocksize - 1)) == 0)))
4671                 goto out;
4672
4673         ret = btrfs_delalloc_reserve_space(inode,
4674                         round_down(from, blocksize), blocksize);
4675         if (ret)
4676                 goto out;
4677
4678 again:
4679         page = find_or_create_page(mapping, index, mask);
4680         if (!page) {
4681                 btrfs_delalloc_release_space(inode,
4682                                 round_down(from, blocksize),
4683                                 blocksize);
4684                 ret = -ENOMEM;
4685                 goto out;
4686         }
4687
4688         block_start = round_down(from, blocksize);
4689         block_end = block_start + blocksize - 1;
4690
4691         if (!PageUptodate(page)) {
4692                 ret = btrfs_readpage(NULL, page);
4693                 lock_page(page);
4694                 if (page->mapping != mapping) {
4695                         unlock_page(page);
4696                         put_page(page);
4697                         goto again;
4698                 }
4699                 if (!PageUptodate(page)) {
4700                         ret = -EIO;
4701                         goto out_unlock;
4702                 }
4703         }
4704         wait_on_page_writeback(page);
4705
4706         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4707         set_page_extent_mapped(page);
4708
4709         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4710         if (ordered) {
4711                 unlock_extent_cached(io_tree, block_start, block_end,
4712                                      &cached_state, GFP_NOFS);
4713                 unlock_page(page);
4714                 put_page(page);
4715                 btrfs_start_ordered_extent(inode, ordered, 1);
4716                 btrfs_put_ordered_extent(ordered);
4717                 goto again;
4718         }
4719
4720         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4721                           EXTENT_DIRTY | EXTENT_DELALLOC |
4722                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4723                           0, 0, &cached_state, GFP_NOFS);
4724
4725         ret = btrfs_set_extent_delalloc(inode, block_start, block_end,
4726                                         &cached_state);
4727         if (ret) {
4728                 unlock_extent_cached(io_tree, block_start, block_end,
4729                                      &cached_state, GFP_NOFS);
4730                 goto out_unlock;
4731         }
4732
4733         if (offset != blocksize) {
4734                 if (!len)
4735                         len = blocksize - offset;
4736                 kaddr = kmap(page);
4737                 if (front)
4738                         memset(kaddr + (block_start - page_offset(page)),
4739                                 0, offset);
4740                 else
4741                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4742                                 0, len);
4743                 flush_dcache_page(page);
4744                 kunmap(page);
4745         }
4746         ClearPageChecked(page);
4747         set_page_dirty(page);
4748         unlock_extent_cached(io_tree, block_start, block_end, &cached_state,
4749                              GFP_NOFS);
4750
4751 out_unlock:
4752         if (ret)
4753                 btrfs_delalloc_release_space(inode, block_start,
4754                                              blocksize);
4755         unlock_page(page);
4756         put_page(page);
4757 out:
4758         return ret;
4759 }
4760
4761 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4762                              u64 offset, u64 len)
4763 {
4764         struct btrfs_trans_handle *trans;
4765         int ret;
4766
4767         /*
4768          * Still need to make sure the inode looks like it's been updated so
4769          * that any holes get logged if we fsync.
4770          */
4771         if (btrfs_fs_incompat(root->fs_info, NO_HOLES)) {
4772                 BTRFS_I(inode)->last_trans = root->fs_info->generation;
4773                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4774                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4775                 return 0;
4776         }
4777
4778         /*
4779          * 1 - for the one we're dropping
4780          * 1 - for the one we're adding
4781          * 1 - for updating the inode.
4782          */
4783         trans = btrfs_start_transaction(root, 3);
4784         if (IS_ERR(trans))
4785                 return PTR_ERR(trans);
4786
4787         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4788         if (ret) {
4789                 btrfs_abort_transaction(trans, root, ret);
4790                 btrfs_end_transaction(trans, root);
4791                 return ret;
4792         }
4793
4794         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
4795                                        0, 0, len, 0, len, 0, 0, 0);
4796         if (ret)
4797                 btrfs_abort_transaction(trans, root, ret);
4798         else
4799                 btrfs_update_inode(trans, root, inode);
4800         btrfs_end_transaction(trans, root);
4801         return ret;
4802 }
4803
4804 /*
4805  * This function puts in dummy file extents for the area we're creating a hole
4806  * for.  So if we are truncating this file to a larger size we need to insert
4807  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4808  * the range between oldsize and size
4809  */
4810 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4811 {
4812         struct btrfs_root *root = BTRFS_I(inode)->root;
4813         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4814         struct extent_map *em = NULL;
4815         struct extent_state *cached_state = NULL;
4816         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4817         u64 hole_start = ALIGN(oldsize, root->sectorsize);
4818         u64 block_end = ALIGN(size, root->sectorsize);
4819         u64 last_byte;
4820         u64 cur_offset;
4821         u64 hole_size;
4822         int err = 0;
4823
4824         /*
4825          * If our size started in the middle of a block we need to zero out the
4826          * rest of the block before we expand the i_size, otherwise we could
4827          * expose stale data.
4828          */
4829         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4830         if (err)
4831                 return err;
4832
4833         if (size <= hole_start)
4834                 return 0;
4835
4836         while (1) {
4837                 struct btrfs_ordered_extent *ordered;
4838
4839                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4840                                  &cached_state);
4841                 ordered = btrfs_lookup_ordered_range(inode, hole_start,
4842                                                      block_end - hole_start);
4843                 if (!ordered)
4844                         break;
4845                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4846                                      &cached_state, GFP_NOFS);
4847                 btrfs_start_ordered_extent(inode, ordered, 1);
4848                 btrfs_put_ordered_extent(ordered);
4849         }
4850
4851         cur_offset = hole_start;
4852         while (1) {
4853                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
4854                                 block_end - cur_offset, 0);
4855                 if (IS_ERR(em)) {
4856                         err = PTR_ERR(em);
4857                         em = NULL;
4858                         break;
4859                 }
4860                 last_byte = min(extent_map_end(em), block_end);
4861                 last_byte = ALIGN(last_byte , root->sectorsize);
4862                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4863                         struct extent_map *hole_em;
4864                         hole_size = last_byte - cur_offset;
4865
4866                         err = maybe_insert_hole(root, inode, cur_offset,
4867                                                 hole_size);
4868                         if (err)
4869                                 break;
4870                         btrfs_drop_extent_cache(inode, cur_offset,
4871                                                 cur_offset + hole_size - 1, 0);
4872                         hole_em = alloc_extent_map();
4873                         if (!hole_em) {
4874                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4875                                         &BTRFS_I(inode)->runtime_flags);
4876                                 goto next;
4877                         }
4878                         hole_em->start = cur_offset;
4879                         hole_em->len = hole_size;
4880                         hole_em->orig_start = cur_offset;
4881
4882                         hole_em->block_start = EXTENT_MAP_HOLE;
4883                         hole_em->block_len = 0;
4884                         hole_em->orig_block_len = 0;
4885                         hole_em->ram_bytes = hole_size;
4886                         hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
4887                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4888                         hole_em->generation = root->fs_info->generation;
4889
4890                         while (1) {
4891                                 write_lock(&em_tree->lock);
4892                                 err = add_extent_mapping(em_tree, hole_em, 1);
4893                                 write_unlock(&em_tree->lock);
4894                                 if (err != -EEXIST)
4895                                         break;
4896                                 btrfs_drop_extent_cache(inode, cur_offset,
4897                                                         cur_offset +
4898                                                         hole_size - 1, 0);
4899                         }
4900                         free_extent_map(hole_em);
4901                 }
4902 next:
4903                 free_extent_map(em);
4904                 em = NULL;
4905                 cur_offset = last_byte;
4906                 if (cur_offset >= block_end)
4907                         break;
4908         }
4909         free_extent_map(em);
4910         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4911                              GFP_NOFS);
4912         return err;
4913 }
4914
4915 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4916 {
4917         struct btrfs_root *root = BTRFS_I(inode)->root;
4918         struct btrfs_trans_handle *trans;
4919         loff_t oldsize = i_size_read(inode);
4920         loff_t newsize = attr->ia_size;
4921         int mask = attr->ia_valid;
4922         int ret;
4923
4924         /*
4925          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4926          * special case where we need to update the times despite not having
4927          * these flags set.  For all other operations the VFS set these flags
4928          * explicitly if it wants a timestamp update.
4929          */
4930         if (newsize != oldsize) {
4931                 inode_inc_iversion(inode);
4932                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4933                         inode->i_ctime = inode->i_mtime =
4934                                 current_fs_time(inode->i_sb);
4935         }
4936
4937         if (newsize > oldsize) {
4938                 /*
4939                  * Don't do an expanding truncate while snapshoting is ongoing.
4940                  * This is to ensure the snapshot captures a fully consistent
4941                  * state of this file - if the snapshot captures this expanding
4942                  * truncation, it must capture all writes that happened before
4943                  * this truncation.
4944                  */
4945                 btrfs_wait_for_snapshot_creation(root);
4946                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4947                 if (ret) {
4948                         btrfs_end_write_no_snapshoting(root);
4949                         return ret;
4950                 }
4951
4952                 trans = btrfs_start_transaction(root, 1);
4953                 if (IS_ERR(trans)) {
4954                         btrfs_end_write_no_snapshoting(root);
4955                         return PTR_ERR(trans);
4956                 }
4957
4958                 i_size_write(inode, newsize);
4959                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
4960                 pagecache_isize_extended(inode, oldsize, newsize);
4961                 ret = btrfs_update_inode(trans, root, inode);
4962                 btrfs_end_write_no_snapshoting(root);
4963                 btrfs_end_transaction(trans, root);
4964         } else {
4965
4966                 /*
4967                  * We're truncating a file that used to have good data down to
4968                  * zero. Make sure it gets into the ordered flush list so that
4969                  * any new writes get down to disk quickly.
4970                  */
4971                 if (newsize == 0)
4972                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4973                                 &BTRFS_I(inode)->runtime_flags);
4974
4975                 /*
4976                  * 1 for the orphan item we're going to add
4977                  * 1 for the orphan item deletion.
4978                  */
4979                 trans = btrfs_start_transaction(root, 2);
4980                 if (IS_ERR(trans))
4981                         return PTR_ERR(trans);
4982
4983                 /*
4984                  * We need to do this in case we fail at _any_ point during the
4985                  * actual truncate.  Once we do the truncate_setsize we could
4986                  * invalidate pages which forces any outstanding ordered io to
4987                  * be instantly completed which will give us extents that need
4988                  * to be truncated.  If we fail to get an orphan inode down we
4989                  * could have left over extents that were never meant to live,
4990                  * so we need to guarantee from this point on that everything
4991                  * will be consistent.
4992                  */
4993                 ret = btrfs_orphan_add(trans, inode);
4994                 btrfs_end_transaction(trans, root);
4995                 if (ret)
4996                         return ret;
4997
4998                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
4999                 truncate_setsize(inode, newsize);
5000
5001                 /* Disable nonlocked read DIO to avoid the end less truncate */
5002                 btrfs_inode_block_unlocked_dio(inode);
5003                 inode_dio_wait(inode);
5004                 btrfs_inode_resume_unlocked_dio(inode);
5005
5006                 ret = btrfs_truncate(inode);
5007                 if (ret && inode->i_nlink) {
5008                         int err;
5009
5010                         /*
5011                          * failed to truncate, disk_i_size is only adjusted down
5012                          * as we remove extents, so it should represent the true
5013                          * size of the inode, so reset the in memory size and
5014                          * delete our orphan entry.
5015                          */
5016                         trans = btrfs_join_transaction(root);
5017                         if (IS_ERR(trans)) {
5018                                 btrfs_orphan_del(NULL, inode);
5019                                 return ret;
5020                         }
5021                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5022                         err = btrfs_orphan_del(trans, inode);
5023                         if (err)
5024                                 btrfs_abort_transaction(trans, root, err);
5025                         btrfs_end_transaction(trans, root);
5026                 }
5027         }
5028
5029         return ret;
5030 }
5031
5032 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5033 {
5034         struct inode *inode = d_inode(dentry);
5035         struct btrfs_root *root = BTRFS_I(inode)->root;
5036         int err;
5037
5038         if (btrfs_root_readonly(root))
5039                 return -EROFS;
5040
5041         err = inode_change_ok(inode, attr);
5042         if (err)
5043                 return err;
5044
5045         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5046                 err = btrfs_setsize(inode, attr);
5047                 if (err)
5048                         return err;
5049         }
5050
5051         if (attr->ia_valid) {
5052                 setattr_copy(inode, attr);
5053                 inode_inc_iversion(inode);
5054                 err = btrfs_dirty_inode(inode);
5055
5056                 if (!err && attr->ia_valid & ATTR_MODE)
5057                         err = posix_acl_chmod(inode, inode->i_mode);
5058         }
5059
5060         return err;
5061 }
5062
5063 /*
5064  * While truncating the inode pages during eviction, we get the VFS calling
5065  * btrfs_invalidatepage() against each page of the inode. This is slow because
5066  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5067  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5068  * extent_state structures over and over, wasting lots of time.
5069  *
5070  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5071  * those expensive operations on a per page basis and do only the ordered io
5072  * finishing, while we release here the extent_map and extent_state structures,
5073  * without the excessive merging and splitting.
5074  */
5075 static void evict_inode_truncate_pages(struct inode *inode)
5076 {
5077         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5078         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5079         struct rb_node *node;
5080
5081         ASSERT(inode->i_state & I_FREEING);
5082         truncate_inode_pages_final(&inode->i_data);
5083
5084         write_lock(&map_tree->lock);
5085         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5086                 struct extent_map *em;
5087
5088                 node = rb_first(&map_tree->map);
5089                 em = rb_entry(node, struct extent_map, rb_node);
5090                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5091                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5092                 remove_extent_mapping(map_tree, em);
5093                 free_extent_map(em);
5094                 if (need_resched()) {
5095                         write_unlock(&map_tree->lock);
5096                         cond_resched();
5097                         write_lock(&map_tree->lock);
5098                 }
5099         }
5100         write_unlock(&map_tree->lock);
5101
5102         /*
5103          * Keep looping until we have no more ranges in the io tree.
5104          * We can have ongoing bios started by readpages (called from readahead)
5105          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5106          * still in progress (unlocked the pages in the bio but did not yet
5107          * unlocked the ranges in the io tree). Therefore this means some
5108          * ranges can still be locked and eviction started because before
5109          * submitting those bios, which are executed by a separate task (work
5110          * queue kthread), inode references (inode->i_count) were not taken
5111          * (which would be dropped in the end io callback of each bio).
5112          * Therefore here we effectively end up waiting for those bios and
5113          * anyone else holding locked ranges without having bumped the inode's
5114          * reference count - if we don't do it, when they access the inode's
5115          * io_tree to unlock a range it may be too late, leading to an
5116          * use-after-free issue.
5117          */
5118         spin_lock(&io_tree->lock);
5119         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5120                 struct extent_state *state;
5121                 struct extent_state *cached_state = NULL;
5122                 u64 start;
5123                 u64 end;
5124
5125                 node = rb_first(&io_tree->state);
5126                 state = rb_entry(node, struct extent_state, rb_node);
5127                 start = state->start;
5128                 end = state->end;
5129                 spin_unlock(&io_tree->lock);
5130
5131                 lock_extent_bits(io_tree, start, end, &cached_state);
5132
5133                 /*
5134                  * If still has DELALLOC flag, the extent didn't reach disk,
5135                  * and its reserved space won't be freed by delayed_ref.
5136                  * So we need to free its reserved space here.
5137                  * (Refer to comment in btrfs_invalidatepage, case 2)
5138                  *
5139                  * Note, end is the bytenr of last byte, so we need + 1 here.
5140                  */
5141                 if (state->state & EXTENT_DELALLOC)
5142                         btrfs_qgroup_free_data(inode, start, end - start + 1);
5143
5144                 clear_extent_bit(io_tree, start, end,
5145                                  EXTENT_LOCKED | EXTENT_DIRTY |
5146                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5147                                  EXTENT_DEFRAG, 1, 1,
5148                                  &cached_state, GFP_NOFS);
5149
5150                 cond_resched();
5151                 spin_lock(&io_tree->lock);
5152         }
5153         spin_unlock(&io_tree->lock);
5154 }
5155
5156 void btrfs_evict_inode(struct inode *inode)
5157 {
5158         struct btrfs_trans_handle *trans;
5159         struct btrfs_root *root = BTRFS_I(inode)->root;
5160         struct btrfs_block_rsv *rsv, *global_rsv;
5161         int steal_from_global = 0;
5162         u64 min_size;
5163         int ret;
5164
5165         trace_btrfs_inode_evict(inode);
5166
5167         if (!root) {
5168                 kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
5169                 return;
5170         }
5171
5172         min_size = btrfs_calc_trunc_metadata_size(root, 1);
5173
5174         evict_inode_truncate_pages(inode);
5175
5176         if (inode->i_nlink &&
5177             ((btrfs_root_refs(&root->root_item) != 0 &&
5178               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5179              btrfs_is_free_space_inode(inode)))
5180                 goto no_delete;
5181
5182         if (is_bad_inode(inode)) {
5183                 btrfs_orphan_del(NULL, inode);
5184                 goto no_delete;
5185         }
5186         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5187         if (!special_file(inode->i_mode))
5188                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5189
5190         btrfs_free_io_failure_record(inode, 0, (u64)-1);
5191
5192         if (root->fs_info->log_root_recovering) {
5193                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5194                                  &BTRFS_I(inode)->runtime_flags));
5195                 goto no_delete;
5196         }
5197
5198         if (inode->i_nlink > 0) {
5199                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5200                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5201                 goto no_delete;
5202         }
5203
5204         ret = btrfs_commit_inode_delayed_inode(inode);
5205         if (ret) {
5206                 btrfs_orphan_del(NULL, inode);
5207                 goto no_delete;
5208         }
5209
5210         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
5211         if (!rsv) {
5212                 btrfs_orphan_del(NULL, inode);
5213                 goto no_delete;
5214         }
5215         rsv->size = min_size;
5216         rsv->failfast = 1;
5217         global_rsv = &root->fs_info->global_block_rsv;
5218
5219         btrfs_i_size_write(inode, 0);
5220
5221         /*
5222          * This is a bit simpler than btrfs_truncate since we've already
5223          * reserved our space for our orphan item in the unlink, so we just
5224          * need to reserve some slack space in case we add bytes and update
5225          * inode item when doing the truncate.
5226          */
5227         while (1) {
5228                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5229                                              BTRFS_RESERVE_FLUSH_LIMIT);
5230
5231                 /*
5232                  * Try and steal from the global reserve since we will
5233                  * likely not use this space anyway, we want to try as
5234                  * hard as possible to get this to work.
5235                  */
5236                 if (ret)
5237                         steal_from_global++;
5238                 else
5239                         steal_from_global = 0;
5240                 ret = 0;
5241
5242                 /*
5243                  * steal_from_global == 0: we reserved stuff, hooray!
5244                  * steal_from_global == 1: we didn't reserve stuff, boo!
5245                  * steal_from_global == 2: we've committed, still not a lot of
5246                  * room but maybe we'll have room in the global reserve this
5247                  * time.
5248                  * steal_from_global == 3: abandon all hope!
5249                  */
5250                 if (steal_from_global > 2) {
5251                         btrfs_warn(root->fs_info,
5252                                 "Could not get space for a delete, will truncate on mount %d",
5253                                 ret);
5254                         btrfs_orphan_del(NULL, inode);
5255                         btrfs_free_block_rsv(root, rsv);
5256                         goto no_delete;
5257                 }
5258
5259                 trans = btrfs_join_transaction(root);
5260                 if (IS_ERR(trans)) {
5261                         btrfs_orphan_del(NULL, inode);
5262                         btrfs_free_block_rsv(root, rsv);
5263                         goto no_delete;
5264                 }
5265
5266                 /*
5267                  * We can't just steal from the global reserve, we need to make
5268                  * sure there is room to do it, if not we need to commit and try
5269                  * again.
5270                  */
5271                 if (steal_from_global) {
5272                         if (!btrfs_check_space_for_delayed_refs(trans, root))
5273                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5274                                                               min_size, 0);
5275                         else
5276                                 ret = -ENOSPC;
5277                 }
5278
5279                 /*
5280                  * Couldn't steal from the global reserve, we have too much
5281                  * pending stuff built up, commit the transaction and try it
5282                  * again.
5283                  */
5284                 if (ret) {
5285                         ret = btrfs_commit_transaction(trans, root);
5286                         if (ret) {
5287                                 btrfs_orphan_del(NULL, inode);
5288                                 btrfs_free_block_rsv(root, rsv);
5289                                 goto no_delete;
5290                         }
5291                         continue;
5292                 } else {
5293                         steal_from_global = 0;
5294                 }
5295
5296                 trans->block_rsv = rsv;
5297
5298                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5299                 if (ret != -ENOSPC && ret != -EAGAIN)
5300                         break;
5301
5302                 trans->block_rsv = &root->fs_info->trans_block_rsv;
5303                 btrfs_end_transaction(trans, root);
5304                 trans = NULL;
5305                 btrfs_btree_balance_dirty(root);
5306         }
5307
5308         btrfs_free_block_rsv(root, rsv);
5309
5310         /*
5311          * Errors here aren't a big deal, it just means we leave orphan items
5312          * in the tree.  They will be cleaned up on the next mount.
5313          */
5314         if (ret == 0) {
5315                 trans->block_rsv = root->orphan_block_rsv;
5316                 btrfs_orphan_del(trans, inode);
5317         } else {
5318                 btrfs_orphan_del(NULL, inode);
5319         }
5320
5321         trans->block_rsv = &root->fs_info->trans_block_rsv;
5322         if (!(root == root->fs_info->tree_root ||
5323               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5324                 btrfs_return_ino(root, btrfs_ino(inode));
5325
5326         btrfs_end_transaction(trans, root);
5327         btrfs_btree_balance_dirty(root);
5328 no_delete:
5329         btrfs_remove_delayed_node(inode);
5330         clear_inode(inode);
5331 }
5332
5333 /*
5334  * this returns the key found in the dir entry in the location pointer.
5335  * If no dir entries were found, location->objectid is 0.
5336  */
5337 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5338                                struct btrfs_key *location)
5339 {
5340         const char *name = dentry->d_name.name;
5341         int namelen = dentry->d_name.len;
5342         struct btrfs_dir_item *di;
5343         struct btrfs_path *path;
5344         struct btrfs_root *root = BTRFS_I(dir)->root;
5345         int ret = 0;
5346
5347         path = btrfs_alloc_path();
5348         if (!path)
5349                 return -ENOMEM;
5350
5351         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(dir), name,
5352                                     namelen, 0);
5353         if (IS_ERR(di))
5354                 ret = PTR_ERR(di);
5355
5356         if (IS_ERR_OR_NULL(di))
5357                 goto out_err;
5358
5359         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5360 out:
5361         btrfs_free_path(path);
5362         return ret;
5363 out_err:
5364         location->objectid = 0;
5365         goto out;
5366 }
5367
5368 /*
5369  * when we hit a tree root in a directory, the btrfs part of the inode
5370  * needs to be changed to reflect the root directory of the tree root.  This
5371  * is kind of like crossing a mount point.
5372  */
5373 static int fixup_tree_root_location(struct btrfs_root *root,
5374                                     struct inode *dir,
5375                                     struct dentry *dentry,
5376                                     struct btrfs_key *location,
5377                                     struct btrfs_root **sub_root)
5378 {
5379         struct btrfs_path *path;
5380         struct btrfs_root *new_root;
5381         struct btrfs_root_ref *ref;
5382         struct extent_buffer *leaf;
5383         struct btrfs_key key;
5384         int ret;
5385         int err = 0;
5386
5387         path = btrfs_alloc_path();
5388         if (!path) {
5389                 err = -ENOMEM;
5390                 goto out;
5391         }
5392
5393         err = -ENOENT;
5394         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5395         key.type = BTRFS_ROOT_REF_KEY;
5396         key.offset = location->objectid;
5397
5398         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, path,
5399                                 0, 0);
5400         if (ret) {
5401                 if (ret < 0)
5402                         err = ret;
5403                 goto out;
5404         }
5405
5406         leaf = path->nodes[0];
5407         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5408         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(dir) ||
5409             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5410                 goto out;
5411
5412         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5413                                    (unsigned long)(ref + 1),
5414                                    dentry->d_name.len);
5415         if (ret)
5416                 goto out;
5417
5418         btrfs_release_path(path);
5419
5420         new_root = btrfs_read_fs_root_no_name(root->fs_info, location);
5421         if (IS_ERR(new_root)) {
5422                 err = PTR_ERR(new_root);
5423                 goto out;
5424         }
5425
5426         *sub_root = new_root;
5427         location->objectid = btrfs_root_dirid(&new_root->root_item);
5428         location->type = BTRFS_INODE_ITEM_KEY;
5429         location->offset = 0;
5430         err = 0;
5431 out:
5432         btrfs_free_path(path);
5433         return err;
5434 }
5435
5436 static void inode_tree_add(struct inode *inode)
5437 {
5438         struct btrfs_root *root = BTRFS_I(inode)->root;
5439         struct btrfs_inode *entry;
5440         struct rb_node **p;
5441         struct rb_node *parent;
5442         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5443         u64 ino = btrfs_ino(inode);
5444
5445         if (inode_unhashed(inode))
5446                 return;
5447         parent = NULL;
5448         spin_lock(&root->inode_lock);
5449         p = &root->inode_tree.rb_node;
5450         while (*p) {
5451                 parent = *p;
5452                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5453
5454                 if (ino < btrfs_ino(&entry->vfs_inode))
5455                         p = &parent->rb_left;
5456                 else if (ino > btrfs_ino(&entry->vfs_inode))
5457                         p = &parent->rb_right;
5458                 else {
5459                         WARN_ON(!(entry->vfs_inode.i_state &
5460                                   (I_WILL_FREE | I_FREEING)));
5461                         rb_replace_node(parent, new, &root->inode_tree);
5462                         RB_CLEAR_NODE(parent);
5463                         spin_unlock(&root->inode_lock);
5464                         return;
5465                 }
5466         }
5467         rb_link_node(new, parent, p);
5468         rb_insert_color(new, &root->inode_tree);
5469         spin_unlock(&root->inode_lock);
5470 }
5471
5472 static void inode_tree_del(struct inode *inode)
5473 {
5474         struct btrfs_root *root = BTRFS_I(inode)->root;
5475         int empty = 0;
5476
5477         spin_lock(&root->inode_lock);
5478         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5479                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5480                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5481                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5482         }
5483         spin_unlock(&root->inode_lock);
5484
5485         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5486                 synchronize_srcu(&root->fs_info->subvol_srcu);
5487                 spin_lock(&root->inode_lock);
5488                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5489                 spin_unlock(&root->inode_lock);
5490                 if (empty)
5491                         btrfs_add_dead_root(root);
5492         }
5493 }
5494
5495 void btrfs_invalidate_inodes(struct btrfs_root *root)
5496 {
5497         struct rb_node *node;
5498         struct rb_node *prev;
5499         struct btrfs_inode *entry;
5500         struct inode *inode;
5501         u64 objectid = 0;
5502
5503         if (!test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
5504                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5505
5506         spin_lock(&root->inode_lock);
5507 again:
5508         node = root->inode_tree.rb_node;
5509         prev = NULL;
5510         while (node) {
5511                 prev = node;
5512                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5513
5514                 if (objectid < btrfs_ino(&entry->vfs_inode))
5515                         node = node->rb_left;
5516                 else if (objectid > btrfs_ino(&entry->vfs_inode))
5517                         node = node->rb_right;
5518                 else
5519                         break;
5520         }
5521         if (!node) {
5522                 while (prev) {
5523                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5524                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5525                                 node = prev;
5526                                 break;
5527                         }
5528                         prev = rb_next(prev);
5529                 }
5530         }
5531         while (node) {
5532                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5533                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5534                 inode = igrab(&entry->vfs_inode);
5535                 if (inode) {
5536                         spin_unlock(&root->inode_lock);
5537                         if (atomic_read(&inode->i_count) > 1)
5538                                 d_prune_aliases(inode);
5539                         /*
5540                          * btrfs_drop_inode will have it removed from
5541                          * the inode cache when its usage count
5542                          * hits zero.
5543                          */
5544                         iput(inode);
5545                         cond_resched();
5546                         spin_lock(&root->inode_lock);
5547                         goto again;
5548                 }
5549
5550                 if (cond_resched_lock(&root->inode_lock))
5551                         goto again;
5552
5553                 node = rb_next(node);
5554         }
5555         spin_unlock(&root->inode_lock);
5556 }
5557
5558 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5559 {
5560         struct btrfs_iget_args *args = p;
5561         inode->i_ino = args->location->objectid;
5562         memcpy(&BTRFS_I(inode)->location, args->location,
5563                sizeof(*args->location));
5564         BTRFS_I(inode)->root = args->root;
5565         return 0;
5566 }
5567
5568 static int btrfs_find_actor(struct inode *inode, void *opaque)
5569 {
5570         struct btrfs_iget_args *args = opaque;
5571         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5572                 args->root == BTRFS_I(inode)->root;
5573 }
5574
5575 static struct inode *btrfs_iget_locked(struct super_block *s,
5576                                        struct btrfs_key *location,
5577                                        struct btrfs_root *root)
5578 {
5579         struct inode *inode;
5580         struct btrfs_iget_args args;
5581         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5582
5583         args.location = location;
5584         args.root = root;
5585
5586         inode = iget5_locked(s, hashval, btrfs_find_actor,
5587                              btrfs_init_locked_inode,
5588                              (void *)&args);
5589         return inode;
5590 }
5591
5592 /* Get an inode object given its location and corresponding root.
5593  * Returns in *is_new if the inode was read from disk
5594  */
5595 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5596                          struct btrfs_root *root, int *new)
5597 {
5598         struct inode *inode;
5599
5600         inode = btrfs_iget_locked(s, location, root);
5601         if (!inode)
5602                 return ERR_PTR(-ENOMEM);
5603
5604         if (inode->i_state & I_NEW) {
5605                 btrfs_read_locked_inode(inode);
5606                 if (!is_bad_inode(inode)) {
5607                         inode_tree_add(inode);
5608                         unlock_new_inode(inode);
5609                         if (new)
5610                                 *new = 1;
5611                 } else {
5612                         unlock_new_inode(inode);
5613                         iput(inode);
5614                         inode = ERR_PTR(-ESTALE);
5615                 }
5616         }
5617
5618         return inode;
5619 }
5620
5621 static struct inode *new_simple_dir(struct super_block *s,
5622                                     struct btrfs_key *key,
5623                                     struct btrfs_root *root)
5624 {
5625         struct inode *inode = new_inode(s);
5626
5627         if (!inode)
5628                 return ERR_PTR(-ENOMEM);
5629
5630         BTRFS_I(inode)->root = root;
5631         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5632         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5633
5634         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5635         inode->i_op = &btrfs_dir_ro_inode_operations;
5636         inode->i_fop = &simple_dir_operations;
5637         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5638         inode->i_mtime = current_fs_time(inode->i_sb);
5639         inode->i_atime = inode->i_mtime;
5640         inode->i_ctime = inode->i_mtime;
5641         BTRFS_I(inode)->i_otime = inode->i_mtime;
5642
5643         return inode;
5644 }
5645
5646 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5647 {
5648         struct inode *inode;
5649         struct btrfs_root *root = BTRFS_I(dir)->root;
5650         struct btrfs_root *sub_root = root;
5651         struct btrfs_key location;
5652         int index;
5653         int ret = 0;
5654
5655         if (dentry->d_name.len > BTRFS_NAME_LEN)
5656                 return ERR_PTR(-ENAMETOOLONG);
5657
5658         ret = btrfs_inode_by_name(dir, dentry, &location);
5659         if (ret < 0)
5660                 return ERR_PTR(ret);
5661
5662         if (location.objectid == 0)
5663                 return ERR_PTR(-ENOENT);
5664
5665         if (location.type == BTRFS_INODE_ITEM_KEY) {
5666                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5667                 return inode;
5668         }
5669
5670         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5671
5672         index = srcu_read_lock(&root->fs_info->subvol_srcu);
5673         ret = fixup_tree_root_location(root, dir, dentry,
5674                                        &location, &sub_root);
5675         if (ret < 0) {
5676                 if (ret != -ENOENT)
5677                         inode = ERR_PTR(ret);
5678                 else
5679                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5680         } else {
5681                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5682         }
5683         srcu_read_unlock(&root->fs_info->subvol_srcu, index);
5684
5685         if (!IS_ERR(inode) && root != sub_root) {
5686                 down_read(&root->fs_info->cleanup_work_sem);
5687                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5688                         ret = btrfs_orphan_cleanup(sub_root);
5689                 up_read(&root->fs_info->cleanup_work_sem);
5690                 if (ret) {
5691                         iput(inode);
5692                         inode = ERR_PTR(ret);
5693                 }
5694         }
5695
5696         return inode;
5697 }
5698
5699 static int btrfs_dentry_delete(const struct dentry *dentry)
5700 {
5701         struct btrfs_root *root;
5702         struct inode *inode = d_inode(dentry);
5703
5704         if (!inode && !IS_ROOT(dentry))
5705                 inode = d_inode(dentry->d_parent);
5706
5707         if (inode) {
5708                 root = BTRFS_I(inode)->root;
5709                 if (btrfs_root_refs(&root->root_item) == 0)
5710                         return 1;
5711
5712                 if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5713                         return 1;
5714         }
5715         return 0;
5716 }
5717
5718 static void btrfs_dentry_release(struct dentry *dentry)
5719 {
5720         kfree(dentry->d_fsdata);
5721 }
5722
5723 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5724                                    unsigned int flags)
5725 {
5726         struct inode *inode;
5727
5728         inode = btrfs_lookup_dentry(dir, dentry);
5729         if (IS_ERR(inode)) {
5730                 if (PTR_ERR(inode) == -ENOENT)
5731                         inode = NULL;
5732                 else
5733                         return ERR_CAST(inode);
5734         }
5735
5736         return d_splice_alias(inode, dentry);
5737 }
5738
5739 unsigned char btrfs_filetype_table[] = {
5740         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5741 };
5742
5743 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5744 {
5745         struct inode *inode = file_inode(file);
5746         struct btrfs_root *root = BTRFS_I(inode)->root;
5747         struct btrfs_item *item;
5748         struct btrfs_dir_item *di;
5749         struct btrfs_key key;
5750         struct btrfs_key found_key;
5751         struct btrfs_path *path;
5752         struct list_head ins_list;
5753         struct list_head del_list;
5754         int ret;
5755         struct extent_buffer *leaf;
5756         int slot;
5757         unsigned char d_type;
5758         int over = 0;
5759         u32 di_cur;
5760         u32 di_total;
5761         u32 di_len;
5762         int key_type = BTRFS_DIR_INDEX_KEY;
5763         char tmp_name[32];
5764         char *name_ptr;
5765         int name_len;
5766         int is_curr = 0;        /* ctx->pos points to the current index? */
5767         bool emitted;
5768         bool put = false;
5769
5770         /* FIXME, use a real flag for deciding about the key type */
5771         if (root->fs_info->tree_root == root)
5772                 key_type = BTRFS_DIR_ITEM_KEY;
5773
5774         if (!dir_emit_dots(file, ctx))
5775                 return 0;
5776
5777         path = btrfs_alloc_path();
5778         if (!path)
5779                 return -ENOMEM;
5780
5781         path->reada = READA_FORWARD;
5782
5783         if (key_type == BTRFS_DIR_INDEX_KEY) {
5784                 INIT_LIST_HEAD(&ins_list);
5785                 INIT_LIST_HEAD(&del_list);
5786                 put = btrfs_readdir_get_delayed_items(inode, &ins_list,
5787                                                       &del_list);
5788         }
5789
5790         key.type = key_type;
5791         key.offset = ctx->pos;
5792         key.objectid = btrfs_ino(inode);
5793
5794         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5795         if (ret < 0)
5796                 goto err;
5797
5798         emitted = false;
5799         while (1) {
5800                 leaf = path->nodes[0];
5801                 slot = path->slots[0];
5802                 if (slot >= btrfs_header_nritems(leaf)) {
5803                         ret = btrfs_next_leaf(root, path);
5804                         if (ret < 0)
5805                                 goto err;
5806                         else if (ret > 0)
5807                                 break;
5808                         continue;
5809                 }
5810
5811                 item = btrfs_item_nr(slot);
5812                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5813
5814                 if (found_key.objectid != key.objectid)
5815                         break;
5816                 if (found_key.type != key_type)
5817                         break;
5818                 if (found_key.offset < ctx->pos)
5819                         goto next;
5820                 if (key_type == BTRFS_DIR_INDEX_KEY &&
5821                     btrfs_should_delete_dir_index(&del_list,
5822                                                   found_key.offset))
5823                         goto next;
5824
5825                 ctx->pos = found_key.offset;
5826                 is_curr = 1;
5827
5828                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5829                 di_cur = 0;
5830                 di_total = btrfs_item_size(leaf, item);
5831
5832                 while (di_cur < di_total) {
5833                         struct btrfs_key location;
5834
5835                         if (verify_dir_item(root, leaf, di))
5836                                 break;
5837
5838                         name_len = btrfs_dir_name_len(leaf, di);
5839                         if (name_len <= sizeof(tmp_name)) {
5840                                 name_ptr = tmp_name;
5841                         } else {
5842                                 name_ptr = kmalloc(name_len, GFP_KERNEL);
5843                                 if (!name_ptr) {
5844                                         ret = -ENOMEM;
5845                                         goto err;
5846                                 }
5847                         }
5848                         read_extent_buffer(leaf, name_ptr,
5849                                            (unsigned long)(di + 1), name_len);
5850
5851                         d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5852                         btrfs_dir_item_key_to_cpu(leaf, di, &location);
5853
5854
5855                         /* is this a reference to our own snapshot? If so
5856                          * skip it.
5857                          *
5858                          * In contrast to old kernels, we insert the snapshot's
5859                          * dir item and dir index after it has been created, so
5860                          * we won't find a reference to our own snapshot. We
5861                          * still keep the following code for backward
5862                          * compatibility.
5863                          */
5864                         if (location.type == BTRFS_ROOT_ITEM_KEY &&
5865                             location.objectid == root->root_key.objectid) {
5866                                 over = 0;
5867                                 goto skip;
5868                         }
5869                         over = !dir_emit(ctx, name_ptr, name_len,
5870                                        location.objectid, d_type);
5871
5872 skip:
5873                         if (name_ptr != tmp_name)
5874                                 kfree(name_ptr);
5875
5876                         if (over)
5877                                 goto nopos;
5878                         emitted = true;
5879                         di_len = btrfs_dir_name_len(leaf, di) +
5880                                  btrfs_dir_data_len(leaf, di) + sizeof(*di);
5881                         di_cur += di_len;
5882                         di = (struct btrfs_dir_item *)((char *)di + di_len);
5883                 }
5884 next:
5885                 path->slots[0]++;
5886         }
5887
5888         if (key_type == BTRFS_DIR_INDEX_KEY) {
5889                 if (is_curr)
5890                         ctx->pos++;
5891                 ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list, &emitted);
5892                 if (ret)
5893                         goto nopos;
5894         }
5895
5896         /*
5897          * If we haven't emitted any dir entry, we must not touch ctx->pos as
5898          * it was was set to the termination value in previous call. We assume
5899          * that "." and ".." were emitted if we reach this point and set the
5900          * termination value as well for an empty directory.
5901          */
5902         if (ctx->pos > 2 && !emitted)
5903                 goto nopos;
5904
5905         /* Reached end of directory/root. Bump pos past the last item. */
5906         ctx->pos++;
5907
5908         /*
5909          * Stop new entries from being returned after we return the last
5910          * entry.
5911          *
5912          * New directory entries are assigned a strictly increasing
5913          * offset.  This means that new entries created during readdir
5914          * are *guaranteed* to be seen in the future by that readdir.
5915          * This has broken buggy programs which operate on names as
5916          * they're returned by readdir.  Until we re-use freed offsets
5917          * we have this hack to stop new entries from being returned
5918          * under the assumption that they'll never reach this huge
5919          * offset.
5920          *
5921          * This is being careful not to overflow 32bit loff_t unless the
5922          * last entry requires it because doing so has broken 32bit apps
5923          * in the past.
5924          */
5925         if (key_type == BTRFS_DIR_INDEX_KEY) {
5926                 if (ctx->pos >= INT_MAX)
5927                         ctx->pos = LLONG_MAX;
5928                 else
5929                         ctx->pos = INT_MAX;
5930         }
5931 nopos:
5932         ret = 0;
5933 err:
5934         if (put)
5935                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
5936         btrfs_free_path(path);
5937         return ret;
5938 }
5939
5940 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5941 {
5942         struct btrfs_root *root = BTRFS_I(inode)->root;
5943         struct btrfs_trans_handle *trans;
5944         int ret = 0;
5945         bool nolock = false;
5946
5947         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5948                 return 0;
5949
5950         if (btrfs_fs_closing(root->fs_info) && btrfs_is_free_space_inode(inode))
5951                 nolock = true;
5952
5953         if (wbc->sync_mode == WB_SYNC_ALL) {
5954                 if (nolock)
5955                         trans = btrfs_join_transaction_nolock(root);
5956                 else
5957                         trans = btrfs_join_transaction(root);
5958                 if (IS_ERR(trans))
5959                         return PTR_ERR(trans);
5960                 ret = btrfs_commit_transaction(trans, root);
5961         }
5962         return ret;
5963 }
5964
5965 /*
5966  * This is somewhat expensive, updating the tree every time the
5967  * inode changes.  But, it is most likely to find the inode in cache.
5968  * FIXME, needs more benchmarking...there are no reasons other than performance
5969  * to keep or drop this code.
5970  */
5971 static int btrfs_dirty_inode(struct inode *inode)
5972 {
5973         struct btrfs_root *root = BTRFS_I(inode)->root;
5974         struct btrfs_trans_handle *trans;
5975         int ret;
5976
5977         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5978                 return 0;
5979
5980         trans = btrfs_join_transaction(root);
5981         if (IS_ERR(trans))
5982                 return PTR_ERR(trans);
5983
5984         ret = btrfs_update_inode(trans, root, inode);
5985         if (ret && ret == -ENOSPC) {
5986                 /* whoops, lets try again with the full transaction */
5987                 btrfs_end_transaction(trans, root);
5988                 trans = btrfs_start_transaction(root, 1);
5989                 if (IS_ERR(trans))
5990                         return PTR_ERR(trans);
5991
5992                 ret = btrfs_update_inode(trans, root, inode);
5993         }
5994         btrfs_end_transaction(trans, root);
5995         if (BTRFS_I(inode)->delayed_node)
5996                 btrfs_balance_delayed_items(root);
5997
5998         return ret;
5999 }
6000
6001 /*
6002  * This is a copy of file_update_time.  We need this so we can return error on
6003  * ENOSPC for updating the inode in the case of file write and mmap writes.
6004  */
6005 static int btrfs_update_time(struct inode *inode, struct timespec *now,
6006                              int flags)
6007 {
6008         struct btrfs_root *root = BTRFS_I(inode)->root;
6009
6010         if (btrfs_root_readonly(root))
6011                 return -EROFS;
6012
6013         if (flags & S_VERSION)
6014                 inode_inc_iversion(inode);
6015         if (flags & S_CTIME)
6016                 inode->i_ctime = *now;
6017         if (flags & S_MTIME)
6018                 inode->i_mtime = *now;
6019         if (flags & S_ATIME)
6020                 inode->i_atime = *now;
6021         return btrfs_dirty_inode(inode);
6022 }
6023
6024 /*
6025  * find the highest existing sequence number in a directory
6026  * and then set the in-memory index_cnt variable to reflect
6027  * free sequence numbers
6028  */
6029 static int btrfs_set_inode_index_count(struct inode *inode)
6030 {
6031         struct btrfs_root *root = BTRFS_I(inode)->root;
6032         struct btrfs_key key, found_key;
6033         struct btrfs_path *path;
6034         struct extent_buffer *leaf;
6035         int ret;
6036
6037         key.objectid = btrfs_ino(inode);
6038         key.type = BTRFS_DIR_INDEX_KEY;
6039         key.offset = (u64)-1;
6040
6041         path = btrfs_alloc_path();
6042         if (!path)
6043                 return -ENOMEM;
6044
6045         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6046         if (ret < 0)
6047                 goto out;
6048         /* FIXME: we should be able to handle this */
6049         if (ret == 0)
6050                 goto out;
6051         ret = 0;
6052
6053         /*
6054          * MAGIC NUMBER EXPLANATION:
6055          * since we search a directory based on f_pos we have to start at 2
6056          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6057          * else has to start at 2
6058          */
6059         if (path->slots[0] == 0) {
6060                 BTRFS_I(inode)->index_cnt = 2;
6061                 goto out;
6062         }
6063
6064         path->slots[0]--;
6065
6066         leaf = path->nodes[0];
6067         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6068
6069         if (found_key.objectid != btrfs_ino(inode) ||
6070             found_key.type != BTRFS_DIR_INDEX_KEY) {
6071                 BTRFS_I(inode)->index_cnt = 2;
6072                 goto out;
6073         }
6074
6075         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
6076 out:
6077         btrfs_free_path(path);
6078         return ret;
6079 }
6080
6081 /*
6082  * helper to find a free sequence number in a given directory.  This current
6083  * code is very simple, later versions will do smarter things in the btree
6084  */
6085 int btrfs_set_inode_index(struct inode *dir, u64 *index)
6086 {
6087         int ret = 0;
6088
6089         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
6090                 ret = btrfs_inode_delayed_dir_index_count(dir);
6091                 if (ret) {
6092                         ret = btrfs_set_inode_index_count(dir);
6093                         if (ret)
6094                                 return ret;
6095                 }
6096         }
6097
6098         *index = BTRFS_I(dir)->index_cnt;
6099         BTRFS_I(dir)->index_cnt++;
6100
6101         return ret;
6102 }
6103
6104 static int btrfs_insert_inode_locked(struct inode *inode)
6105 {
6106         struct btrfs_iget_args args;
6107         args.location = &BTRFS_I(inode)->location;
6108         args.root = BTRFS_I(inode)->root;
6109
6110         return insert_inode_locked4(inode,
6111                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6112                    btrfs_find_actor, &args);
6113 }
6114
6115 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6116                                      struct btrfs_root *root,
6117                                      struct inode *dir,
6118                                      const char *name, int name_len,
6119                                      u64 ref_objectid, u64 objectid,
6120                                      umode_t mode, u64 *index)
6121 {
6122         struct inode *inode;
6123         struct btrfs_inode_item *inode_item;
6124         struct btrfs_key *location;
6125         struct btrfs_path *path;
6126         struct btrfs_inode_ref *ref;
6127         struct btrfs_key key[2];
6128         u32 sizes[2];
6129         int nitems = name ? 2 : 1;
6130         unsigned long ptr;
6131         int ret;
6132
6133         path = btrfs_alloc_path();
6134         if (!path)
6135                 return ERR_PTR(-ENOMEM);
6136
6137         inode = new_inode(root->fs_info->sb);
6138         if (!inode) {
6139                 btrfs_free_path(path);
6140                 return ERR_PTR(-ENOMEM);
6141         }
6142
6143         /*
6144          * O_TMPFILE, set link count to 0, so that after this point,
6145          * we fill in an inode item with the correct link count.
6146          */
6147         if (!name)
6148                 set_nlink(inode, 0);
6149
6150         /*
6151          * we have to initialize this early, so we can reclaim the inode
6152          * number if we fail afterwards in this function.
6153          */
6154         inode->i_ino = objectid;
6155
6156         if (dir && name) {
6157                 trace_btrfs_inode_request(dir);
6158
6159                 ret = btrfs_set_inode_index(dir, index);
6160                 if (ret) {
6161                         btrfs_free_path(path);
6162                         iput(inode);
6163                         return ERR_PTR(ret);
6164                 }
6165         } else if (dir) {
6166                 *index = 0;
6167         }
6168         /*
6169          * index_cnt is ignored for everything but a dir,
6170          * btrfs_get_inode_index_count has an explanation for the magic
6171          * number
6172          */
6173         BTRFS_I(inode)->index_cnt = 2;
6174         BTRFS_I(inode)->dir_index = *index;
6175         BTRFS_I(inode)->root = root;
6176         BTRFS_I(inode)->generation = trans->transid;
6177         inode->i_generation = BTRFS_I(inode)->generation;
6178
6179         /*
6180          * We could have gotten an inode number from somebody who was fsynced
6181          * and then removed in this same transaction, so let's just set full
6182          * sync since it will be a full sync anyway and this will blow away the
6183          * old info in the log.
6184          */
6185         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6186
6187         key[0].objectid = objectid;
6188         key[0].type = BTRFS_INODE_ITEM_KEY;
6189         key[0].offset = 0;
6190
6191         sizes[0] = sizeof(struct btrfs_inode_item);
6192
6193         if (name) {
6194                 /*
6195                  * Start new inodes with an inode_ref. This is slightly more
6196                  * efficient for small numbers of hard links since they will
6197                  * be packed into one item. Extended refs will kick in if we
6198                  * add more hard links than can fit in the ref item.
6199                  */
6200                 key[1].objectid = objectid;
6201                 key[1].type = BTRFS_INODE_REF_KEY;
6202                 key[1].offset = ref_objectid;
6203
6204                 sizes[1] = name_len + sizeof(*ref);
6205         }
6206
6207         location = &BTRFS_I(inode)->location;
6208         location->objectid = objectid;
6209         location->offset = 0;
6210         location->type = BTRFS_INODE_ITEM_KEY;
6211
6212         ret = btrfs_insert_inode_locked(inode);
6213         if (ret < 0)
6214                 goto fail;
6215
6216         path->leave_spinning = 1;
6217         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6218         if (ret != 0)
6219                 goto fail_unlock;
6220
6221         inode_init_owner(inode, dir, mode);
6222         inode_set_bytes(inode, 0);
6223
6224         inode->i_mtime = current_fs_time(inode->i_sb);
6225         inode->i_atime = inode->i_mtime;
6226         inode->i_ctime = inode->i_mtime;
6227         BTRFS_I(inode)->i_otime = inode->i_mtime;
6228
6229         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6230                                   struct btrfs_inode_item);
6231         memset_extent_buffer(path->nodes[0], 0, (unsigned long)inode_item,
6232                              sizeof(*inode_item));
6233         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6234
6235         if (name) {
6236                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6237                                      struct btrfs_inode_ref);
6238                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6239                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6240                 ptr = (unsigned long)(ref + 1);
6241                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6242         }
6243
6244         btrfs_mark_buffer_dirty(path->nodes[0]);
6245         btrfs_free_path(path);
6246
6247         btrfs_inherit_iflags(inode, dir);
6248
6249         if (S_ISREG(mode)) {
6250                 if (btrfs_test_opt(root, NODATASUM))
6251                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6252                 if (btrfs_test_opt(root, NODATACOW))
6253                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6254                                 BTRFS_INODE_NODATASUM;
6255         }
6256
6257         inode_tree_add(inode);
6258
6259         trace_btrfs_inode_new(inode);
6260         btrfs_set_inode_last_trans(trans, inode);
6261
6262         btrfs_update_root_times(trans, root);
6263
6264         ret = btrfs_inode_inherit_props(trans, inode, dir);
6265         if (ret)
6266                 btrfs_err(root->fs_info,
6267                           "error inheriting props for ino %llu (root %llu): %d",
6268                           btrfs_ino(inode), root->root_key.objectid, ret);
6269
6270         return inode;
6271
6272 fail_unlock:
6273         unlock_new_inode(inode);
6274 fail:
6275         if (dir && name)
6276                 BTRFS_I(dir)->index_cnt--;
6277         btrfs_free_path(path);
6278         iput(inode);
6279         return ERR_PTR(ret);
6280 }
6281
6282 static inline u8 btrfs_inode_type(struct inode *inode)
6283 {
6284         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6285 }
6286
6287 /*
6288  * utility function to add 'inode' into 'parent_inode' with
6289  * a give name and a given sequence number.
6290  * if 'add_backref' is true, also insert a backref from the
6291  * inode to the parent directory.
6292  */
6293 int btrfs_add_link(struct btrfs_trans_handle *trans,
6294                    struct inode *parent_inode, struct inode *inode,
6295                    const char *name, int name_len, int add_backref, u64 index)
6296 {
6297         int ret = 0;
6298         struct btrfs_key key;
6299         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
6300         u64 ino = btrfs_ino(inode);
6301         u64 parent_ino = btrfs_ino(parent_inode);
6302
6303         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6304                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
6305         } else {
6306                 key.objectid = ino;
6307                 key.type = BTRFS_INODE_ITEM_KEY;
6308                 key.offset = 0;
6309         }
6310
6311         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6312                 ret = btrfs_add_root_ref(trans, root->fs_info->tree_root,
6313                                          key.objectid, root->root_key.objectid,
6314                                          parent_ino, index, name, name_len);
6315         } else if (add_backref) {
6316                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6317                                              parent_ino, index);
6318         }
6319
6320         /* Nothing to clean up yet */
6321         if (ret)
6322                 return ret;
6323
6324         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6325                                     parent_inode, &key,
6326                                     btrfs_inode_type(inode), index);
6327         if (ret == -EEXIST || ret == -EOVERFLOW)
6328                 goto fail_dir_item;
6329         else if (ret) {
6330                 btrfs_abort_transaction(trans, root, ret);
6331                 return ret;
6332         }
6333
6334         btrfs_i_size_write(parent_inode, parent_inode->i_size +
6335                            name_len * 2);
6336         inode_inc_iversion(parent_inode);
6337         parent_inode->i_mtime = parent_inode->i_ctime =
6338                 current_fs_time(parent_inode->i_sb);
6339         ret = btrfs_update_inode(trans, root, parent_inode);
6340         if (ret)
6341                 btrfs_abort_transaction(trans, root, ret);
6342         return ret;
6343
6344 fail_dir_item:
6345         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6346                 u64 local_index;
6347                 int err;
6348                 err = btrfs_del_root_ref(trans, root->fs_info->tree_root,
6349                                  key.objectid, root->root_key.objectid,
6350                                  parent_ino, &local_index, name, name_len);
6351
6352         } else if (add_backref) {
6353                 u64 local_index;
6354                 int err;
6355
6356                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6357                                           ino, parent_ino, &local_index);
6358         }
6359         return ret;
6360 }
6361
6362 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6363                             struct inode *dir, struct dentry *dentry,
6364                             struct inode *inode, int backref, u64 index)
6365 {
6366         int err = btrfs_add_link(trans, dir, inode,
6367                                  dentry->d_name.name, dentry->d_name.len,
6368                                  backref, index);
6369         if (err > 0)
6370                 err = -EEXIST;
6371         return err;
6372 }
6373
6374 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6375                         umode_t mode, dev_t rdev)
6376 {
6377         struct btrfs_trans_handle *trans;
6378         struct btrfs_root *root = BTRFS_I(dir)->root;
6379         struct inode *inode = NULL;
6380         int err;
6381         int drop_inode = 0;
6382         u64 objectid;
6383         u64 index = 0;
6384
6385         /*
6386          * 2 for inode item and ref
6387          * 2 for dir items
6388          * 1 for xattr if selinux is on
6389          */
6390         trans = btrfs_start_transaction(root, 5);
6391         if (IS_ERR(trans))
6392                 return PTR_ERR(trans);
6393
6394         err = btrfs_find_free_ino(root, &objectid);
6395         if (err)
6396                 goto out_unlock;
6397
6398         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6399                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6400                                 mode, &index);
6401         if (IS_ERR(inode)) {
6402                 err = PTR_ERR(inode);
6403                 goto out_unlock;
6404         }
6405
6406         /*
6407         * If the active LSM wants to access the inode during
6408         * d_instantiate it needs these. Smack checks to see
6409         * if the filesystem supports xattrs by looking at the
6410         * ops vector.
6411         */
6412         inode->i_op = &btrfs_special_inode_operations;
6413         init_special_inode(inode, inode->i_mode, rdev);
6414
6415         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6416         if (err)
6417                 goto out_unlock_inode;
6418
6419         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6420         if (err) {
6421                 goto out_unlock_inode;
6422         } else {
6423                 btrfs_update_inode(trans, root, inode);
6424                 unlock_new_inode(inode);
6425                 d_instantiate(dentry, inode);
6426         }
6427
6428 out_unlock:
6429         btrfs_end_transaction(trans, root);
6430         btrfs_balance_delayed_items(root);
6431         btrfs_btree_balance_dirty(root);
6432         if (drop_inode) {
6433                 inode_dec_link_count(inode);
6434                 iput(inode);
6435         }
6436         return err;
6437
6438 out_unlock_inode:
6439         drop_inode = 1;
6440         unlock_new_inode(inode);
6441         goto out_unlock;
6442
6443 }
6444
6445 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6446                         umode_t mode, bool excl)
6447 {
6448         struct btrfs_trans_handle *trans;
6449         struct btrfs_root *root = BTRFS_I(dir)->root;
6450         struct inode *inode = NULL;
6451         int drop_inode_on_err = 0;
6452         int err;
6453         u64 objectid;
6454         u64 index = 0;
6455
6456         /*
6457          * 2 for inode item and ref
6458          * 2 for dir items
6459          * 1 for xattr if selinux is on
6460          */
6461         trans = btrfs_start_transaction(root, 5);
6462         if (IS_ERR(trans))
6463                 return PTR_ERR(trans);
6464
6465         err = btrfs_find_free_ino(root, &objectid);
6466         if (err)
6467                 goto out_unlock;
6468
6469         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6470                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6471                                 mode, &index);
6472         if (IS_ERR(inode)) {
6473                 err = PTR_ERR(inode);
6474                 goto out_unlock;
6475         }
6476         drop_inode_on_err = 1;
6477         /*
6478         * If the active LSM wants to access the inode during
6479         * d_instantiate it needs these. Smack checks to see
6480         * if the filesystem supports xattrs by looking at the
6481         * ops vector.
6482         */
6483         inode->i_fop = &btrfs_file_operations;
6484         inode->i_op = &btrfs_file_inode_operations;
6485         inode->i_mapping->a_ops = &btrfs_aops;
6486
6487         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6488         if (err)
6489                 goto out_unlock_inode;
6490
6491         err = btrfs_update_inode(trans, root, inode);
6492         if (err)
6493                 goto out_unlock_inode;
6494
6495         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6496         if (err)
6497                 goto out_unlock_inode;
6498
6499         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6500         unlock_new_inode(inode);
6501         d_instantiate(dentry, inode);
6502
6503 out_unlock:
6504         btrfs_end_transaction(trans, root);
6505         if (err && drop_inode_on_err) {
6506                 inode_dec_link_count(inode);
6507                 iput(inode);
6508         }
6509         btrfs_balance_delayed_items(root);
6510         btrfs_btree_balance_dirty(root);
6511         return err;
6512
6513 out_unlock_inode:
6514         unlock_new_inode(inode);
6515         goto out_unlock;
6516
6517 }
6518
6519 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6520                       struct dentry *dentry)
6521 {
6522         struct btrfs_trans_handle *trans = NULL;
6523         struct btrfs_root *root = BTRFS_I(dir)->root;
6524         struct inode *inode = d_inode(old_dentry);
6525         u64 index;
6526         int err;
6527         int drop_inode = 0;
6528
6529         /* do not allow sys_link's with other subvols of the same device */
6530         if (root->objectid != BTRFS_I(inode)->root->objectid)
6531                 return -EXDEV;
6532
6533         if (inode->i_nlink >= BTRFS_LINK_MAX)
6534                 return -EMLINK;
6535
6536         err = btrfs_set_inode_index(dir, &index);
6537         if (err)
6538                 goto fail;
6539
6540         /*
6541          * 2 items for inode and inode ref
6542          * 2 items for dir items
6543          * 1 item for parent inode
6544          */
6545         trans = btrfs_start_transaction(root, 5);
6546         if (IS_ERR(trans)) {
6547                 err = PTR_ERR(trans);
6548                 trans = NULL;
6549                 goto fail;
6550         }
6551
6552         /* There are several dir indexes for this inode, clear the cache. */
6553         BTRFS_I(inode)->dir_index = 0ULL;
6554         inc_nlink(inode);
6555         inode_inc_iversion(inode);
6556         inode->i_ctime = current_fs_time(inode->i_sb);
6557         ihold(inode);
6558         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6559
6560         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
6561
6562         if (err) {
6563                 drop_inode = 1;
6564         } else {
6565                 struct dentry *parent = dentry->d_parent;
6566                 err = btrfs_update_inode(trans, root, inode);
6567                 if (err)
6568                         goto fail;
6569                 if (inode->i_nlink == 1) {
6570                         /*
6571                          * If new hard link count is 1, it's a file created
6572                          * with open(2) O_TMPFILE flag.
6573                          */
6574                         err = btrfs_orphan_del(trans, inode);
6575                         if (err)
6576                                 goto fail;
6577                 }
6578                 d_instantiate(dentry, inode);
6579                 btrfs_log_new_name(trans, inode, NULL, parent);
6580         }
6581
6582         btrfs_balance_delayed_items(root);
6583 fail:
6584         if (trans)
6585                 btrfs_end_transaction(trans, root);
6586         if (drop_inode) {
6587                 inode_dec_link_count(inode);
6588                 iput(inode);
6589         }
6590         btrfs_btree_balance_dirty(root);
6591         return err;
6592 }
6593
6594 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6595 {
6596         struct inode *inode = NULL;
6597         struct btrfs_trans_handle *trans;
6598         struct btrfs_root *root = BTRFS_I(dir)->root;
6599         int err = 0;
6600         int drop_on_err = 0;
6601         u64 objectid = 0;
6602         u64 index = 0;
6603
6604         /*
6605          * 2 items for inode and ref
6606          * 2 items for dir items
6607          * 1 for xattr if selinux is on
6608          */
6609         trans = btrfs_start_transaction(root, 5);
6610         if (IS_ERR(trans))
6611                 return PTR_ERR(trans);
6612
6613         err = btrfs_find_free_ino(root, &objectid);
6614         if (err)
6615                 goto out_fail;
6616
6617         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6618                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6619                                 S_IFDIR | mode, &index);
6620         if (IS_ERR(inode)) {
6621                 err = PTR_ERR(inode);
6622                 goto out_fail;
6623         }
6624
6625         drop_on_err = 1;
6626         /* these must be set before we unlock the inode */
6627         inode->i_op = &btrfs_dir_inode_operations;
6628         inode->i_fop = &btrfs_dir_file_operations;
6629
6630         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6631         if (err)
6632                 goto out_fail_inode;
6633
6634         btrfs_i_size_write(inode, 0);
6635         err = btrfs_update_inode(trans, root, inode);
6636         if (err)
6637                 goto out_fail_inode;
6638
6639         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
6640                              dentry->d_name.len, 0, index);
6641         if (err)
6642                 goto out_fail_inode;
6643
6644         d_instantiate(dentry, inode);
6645         /*
6646          * mkdir is special.  We're unlocking after we call d_instantiate
6647          * to avoid a race with nfsd calling d_instantiate.
6648          */
6649         unlock_new_inode(inode);
6650         drop_on_err = 0;
6651
6652 out_fail:
6653         btrfs_end_transaction(trans, root);
6654         if (drop_on_err) {
6655                 inode_dec_link_count(inode);
6656                 iput(inode);
6657         }
6658         btrfs_balance_delayed_items(root);
6659         btrfs_btree_balance_dirty(root);
6660         return err;
6661
6662 out_fail_inode:
6663         unlock_new_inode(inode);
6664         goto out_fail;
6665 }
6666
6667 /* Find next extent map of a given extent map, caller needs to ensure locks */
6668 static struct extent_map *next_extent_map(struct extent_map *em)
6669 {
6670         struct rb_node *next;
6671
6672         next = rb_next(&em->rb_node);
6673         if (!next)
6674                 return NULL;
6675         return container_of(next, struct extent_map, rb_node);
6676 }
6677
6678 static struct extent_map *prev_extent_map(struct extent_map *em)
6679 {
6680         struct rb_node *prev;
6681
6682         prev = rb_prev(&em->rb_node);
6683         if (!prev)
6684                 return NULL;
6685         return container_of(prev, struct extent_map, rb_node);
6686 }
6687
6688 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6689  * the existing extent is the nearest extent to map_start,
6690  * and an extent that you want to insert, deal with overlap and insert
6691  * the best fitted new extent into the tree.
6692  */
6693 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6694                                 struct extent_map *existing,
6695                                 struct extent_map *em,
6696                                 u64 map_start)
6697 {
6698         struct extent_map *prev;
6699         struct extent_map *next;
6700         u64 start;
6701         u64 end;
6702         u64 start_diff;
6703
6704         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6705
6706         if (existing->start > map_start) {
6707                 next = existing;
6708                 prev = prev_extent_map(next);
6709         } else {
6710                 prev = existing;
6711                 next = next_extent_map(prev);
6712         }
6713
6714         start = prev ? extent_map_end(prev) : em->start;
6715         start = max_t(u64, start, em->start);
6716         end = next ? next->start : extent_map_end(em);
6717         end = min_t(u64, end, extent_map_end(em));
6718         start_diff = start - em->start;
6719         em->start = start;
6720         em->len = end - start;
6721         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6722             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6723                 em->block_start += start_diff;
6724                 em->block_len -= start_diff;
6725         }
6726         return add_extent_mapping(em_tree, em, 0);
6727 }
6728
6729 static noinline int uncompress_inline(struct btrfs_path *path,
6730                                       struct page *page,
6731                                       size_t pg_offset, u64 extent_offset,
6732                                       struct btrfs_file_extent_item *item)
6733 {
6734         int ret;
6735         struct extent_buffer *leaf = path->nodes[0];
6736         char *tmp;
6737         size_t max_size;
6738         unsigned long inline_size;
6739         unsigned long ptr;
6740         int compress_type;
6741
6742         WARN_ON(pg_offset != 0);
6743         compress_type = btrfs_file_extent_compression(leaf, item);
6744         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6745         inline_size = btrfs_file_extent_inline_item_len(leaf,
6746                                         btrfs_item_nr(path->slots[0]));
6747         tmp = kmalloc(inline_size, GFP_NOFS);
6748         if (!tmp)
6749                 return -ENOMEM;
6750         ptr = btrfs_file_extent_inline_start(item);
6751
6752         read_extent_buffer(leaf, tmp, ptr, inline_size);
6753
6754         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6755         ret = btrfs_decompress(compress_type, tmp, page,
6756                                extent_offset, inline_size, max_size);
6757         kfree(tmp);
6758         return ret;
6759 }
6760
6761 /*
6762  * a bit scary, this does extent mapping from logical file offset to the disk.
6763  * the ugly parts come from merging extents from the disk with the in-ram
6764  * representation.  This gets more complex because of the data=ordered code,
6765  * where the in-ram extents might be locked pending data=ordered completion.
6766  *
6767  * This also copies inline extents directly into the page.
6768  */
6769
6770 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
6771                                     size_t pg_offset, u64 start, u64 len,
6772                                     int create)
6773 {
6774         int ret;
6775         int err = 0;
6776         u64 extent_start = 0;
6777         u64 extent_end = 0;
6778         u64 objectid = btrfs_ino(inode);
6779         u32 found_type;
6780         struct btrfs_path *path = NULL;
6781         struct btrfs_root *root = BTRFS_I(inode)->root;
6782         struct btrfs_file_extent_item *item;
6783         struct extent_buffer *leaf;
6784         struct btrfs_key found_key;
6785         struct extent_map *em = NULL;
6786         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
6787         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6788         struct btrfs_trans_handle *trans = NULL;
6789         const bool new_inline = !page || create;
6790
6791 again:
6792         read_lock(&em_tree->lock);
6793         em = lookup_extent_mapping(em_tree, start, len);
6794         if (em)
6795                 em->bdev = root->fs_info->fs_devices->latest_bdev;
6796         read_unlock(&em_tree->lock);
6797
6798         if (em) {
6799                 if (em->start > start || em->start + em->len <= start)
6800                         free_extent_map(em);
6801                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6802                         free_extent_map(em);
6803                 else
6804                         goto out;
6805         }
6806         em = alloc_extent_map();
6807         if (!em) {
6808                 err = -ENOMEM;
6809                 goto out;
6810         }
6811         em->bdev = root->fs_info->fs_devices->latest_bdev;
6812         em->start = EXTENT_MAP_HOLE;
6813         em->orig_start = EXTENT_MAP_HOLE;
6814         em->len = (u64)-1;
6815         em->block_len = (u64)-1;
6816
6817         if (!path) {
6818                 path = btrfs_alloc_path();
6819                 if (!path) {
6820                         err = -ENOMEM;
6821                         goto out;
6822                 }
6823                 /*
6824                  * Chances are we'll be called again, so go ahead and do
6825                  * readahead
6826                  */
6827                 path->reada = READA_FORWARD;
6828         }
6829
6830         ret = btrfs_lookup_file_extent(trans, root, path,
6831                                        objectid, start, trans != NULL);
6832         if (ret < 0) {
6833                 err = ret;
6834                 goto out;
6835         }
6836
6837         if (ret != 0) {
6838                 if (path->slots[0] == 0)
6839                         goto not_found;
6840                 path->slots[0]--;
6841         }
6842
6843         leaf = path->nodes[0];
6844         item = btrfs_item_ptr(leaf, path->slots[0],
6845                               struct btrfs_file_extent_item);
6846         /* are we inside the extent that was found? */
6847         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6848         found_type = found_key.type;
6849         if (found_key.objectid != objectid ||
6850             found_type != BTRFS_EXTENT_DATA_KEY) {
6851                 /*
6852                  * If we backup past the first extent we want to move forward
6853                  * and see if there is an extent in front of us, otherwise we'll
6854                  * say there is a hole for our whole search range which can
6855                  * cause problems.
6856                  */
6857                 extent_end = start;
6858                 goto next;
6859         }
6860
6861         found_type = btrfs_file_extent_type(leaf, item);
6862         extent_start = found_key.offset;
6863         if (found_type == BTRFS_FILE_EXTENT_REG ||
6864             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6865                 extent_end = extent_start +
6866                        btrfs_file_extent_num_bytes(leaf, item);
6867         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6868                 size_t size;
6869                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6870                 extent_end = ALIGN(extent_start + size, root->sectorsize);
6871         }
6872 next:
6873         if (start >= extent_end) {
6874                 path->slots[0]++;
6875                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6876                         ret = btrfs_next_leaf(root, path);
6877                         if (ret < 0) {
6878                                 err = ret;
6879                                 goto out;
6880                         }
6881                         if (ret > 0)
6882                                 goto not_found;
6883                         leaf = path->nodes[0];
6884                 }
6885                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6886                 if (found_key.objectid != objectid ||
6887                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6888                         goto not_found;
6889                 if (start + len <= found_key.offset)
6890                         goto not_found;
6891                 if (start > found_key.offset)
6892                         goto next;
6893                 em->start = start;
6894                 em->orig_start = start;
6895                 em->len = found_key.offset - start;
6896                 goto not_found_em;
6897         }
6898
6899         btrfs_extent_item_to_extent_map(inode, path, item, new_inline, em);
6900
6901         if (found_type == BTRFS_FILE_EXTENT_REG ||
6902             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6903                 goto insert;
6904         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6905                 unsigned long ptr;
6906                 char *map;
6907                 size_t size;
6908                 size_t extent_offset;
6909                 size_t copy_size;
6910
6911                 if (new_inline)
6912                         goto out;
6913
6914                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6915                 extent_offset = page_offset(page) + pg_offset - extent_start;
6916                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
6917                                   size - extent_offset);
6918                 em->start = extent_start + extent_offset;
6919                 em->len = ALIGN(copy_size, root->sectorsize);
6920                 em->orig_block_len = em->len;
6921                 em->orig_start = em->start;
6922                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6923                 if (create == 0 && !PageUptodate(page)) {
6924                         if (btrfs_file_extent_compression(leaf, item) !=
6925                             BTRFS_COMPRESS_NONE) {
6926                                 ret = uncompress_inline(path, page, pg_offset,
6927                                                         extent_offset, item);
6928                                 if (ret) {
6929                                         err = ret;
6930                                         goto out;
6931                                 }
6932                         } else {
6933                                 map = kmap(page);
6934                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6935                                                    copy_size);
6936                                 if (pg_offset + copy_size < PAGE_SIZE) {
6937                                         memset(map + pg_offset + copy_size, 0,
6938                                                PAGE_SIZE - pg_offset -
6939                                                copy_size);
6940                                 }
6941                                 kunmap(page);
6942                         }
6943                         flush_dcache_page(page);
6944                 } else if (create && PageUptodate(page)) {
6945                         BUG();
6946                         if (!trans) {
6947                                 kunmap(page);
6948                                 free_extent_map(em);
6949                                 em = NULL;
6950
6951                                 btrfs_release_path(path);
6952                                 trans = btrfs_join_transaction(root);
6953
6954                                 if (IS_ERR(trans))
6955                                         return ERR_CAST(trans);
6956                                 goto again;
6957                         }
6958                         map = kmap(page);
6959                         write_extent_buffer(leaf, map + pg_offset, ptr,
6960                                             copy_size);
6961                         kunmap(page);
6962                         btrfs_mark_buffer_dirty(leaf);
6963                 }
6964                 set_extent_uptodate(io_tree, em->start,
6965                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6966                 goto insert;
6967         }
6968 not_found:
6969         em->start = start;
6970         em->orig_start = start;
6971         em->len = len;
6972 not_found_em:
6973         em->block_start = EXTENT_MAP_HOLE;
6974         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
6975 insert:
6976         btrfs_release_path(path);
6977         if (em->start > start || extent_map_end(em) <= start) {
6978                 btrfs_err(root->fs_info, "bad extent! em: [%llu %llu] passed [%llu %llu]",
6979                         em->start, em->len, start, len);
6980                 err = -EIO;
6981                 goto out;
6982         }
6983
6984         err = 0;
6985         write_lock(&em_tree->lock);
6986         ret = add_extent_mapping(em_tree, em, 0);
6987         /* it is possible that someone inserted the extent into the tree
6988          * while we had the lock dropped.  It is also possible that
6989          * an overlapping map exists in the tree
6990          */
6991         if (ret == -EEXIST) {
6992                 struct extent_map *existing;
6993
6994                 ret = 0;
6995
6996                 existing = search_extent_mapping(em_tree, start, len);
6997                 /*
6998                  * existing will always be non-NULL, since there must be
6999                  * extent causing the -EEXIST.
7000                  */
7001                 if (existing->start == em->start &&
7002                     extent_map_end(existing) == extent_map_end(em) &&
7003                     em->block_start == existing->block_start) {
7004                         /*
7005                          * these two extents are the same, it happens
7006                          * with inlines especially
7007                          */
7008                         free_extent_map(em);
7009                         em = existing;
7010                         err = 0;
7011
7012                 } else if (start >= extent_map_end(existing) ||
7013                     start <= existing->start) {
7014                         /*
7015                          * The existing extent map is the one nearest to
7016                          * the [start, start + len) range which overlaps
7017                          */
7018                         err = merge_extent_mapping(em_tree, existing,
7019                                                    em, start);
7020                         free_extent_map(existing);
7021                         if (err) {
7022                                 free_extent_map(em);
7023                                 em = NULL;
7024                         }
7025                 } else {
7026                         free_extent_map(em);
7027                         em = existing;
7028                         err = 0;
7029                 }
7030         }
7031         write_unlock(&em_tree->lock);
7032 out:
7033
7034         trace_btrfs_get_extent(root, em);
7035
7036         btrfs_free_path(path);
7037         if (trans) {
7038                 ret = btrfs_end_transaction(trans, root);
7039                 if (!err)
7040                         err = ret;
7041         }
7042         if (err) {
7043                 free_extent_map(em);
7044                 return ERR_PTR(err);
7045         }
7046         BUG_ON(!em); /* Error is always set */
7047         return em;
7048 }
7049
7050 struct extent_map *btrfs_get_extent_fiemap(struct inode *inode, struct page *page,
7051                                            size_t pg_offset, u64 start, u64 len,
7052                                            int create)
7053 {
7054         struct extent_map *em;
7055         struct extent_map *hole_em = NULL;
7056         u64 range_start = start;
7057         u64 end;
7058         u64 found;
7059         u64 found_end;
7060         int err = 0;
7061
7062         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7063         if (IS_ERR(em))
7064                 return em;
7065         if (em) {
7066                 /*
7067                  * if our em maps to
7068                  * -  a hole or
7069                  * -  a pre-alloc extent,
7070                  * there might actually be delalloc bytes behind it.
7071                  */
7072                 if (em->block_start != EXTENT_MAP_HOLE &&
7073                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7074                         return em;
7075                 else
7076                         hole_em = em;
7077         }
7078
7079         /* check to see if we've wrapped (len == -1 or similar) */
7080         end = start + len;
7081         if (end < start)
7082                 end = (u64)-1;
7083         else
7084                 end -= 1;
7085
7086         em = NULL;
7087
7088         /* ok, we didn't find anything, lets look for delalloc */
7089         found = count_range_bits(&BTRFS_I(inode)->io_tree, &range_start,
7090                                  end, len, EXTENT_DELALLOC, 1);
7091         found_end = range_start + found;
7092         if (found_end < range_start)
7093                 found_end = (u64)-1;
7094
7095         /*
7096          * we didn't find anything useful, return
7097          * the original results from get_extent()
7098          */
7099         if (range_start > end || found_end <= start) {
7100                 em = hole_em;
7101                 hole_em = NULL;
7102                 goto out;
7103         }
7104
7105         /* adjust the range_start to make sure it doesn't
7106          * go backwards from the start they passed in
7107          */
7108         range_start = max(start, range_start);
7109         found = found_end - range_start;
7110
7111         if (found > 0) {
7112                 u64 hole_start = start;
7113                 u64 hole_len = len;
7114
7115                 em = alloc_extent_map();
7116                 if (!em) {
7117                         err = -ENOMEM;
7118                         goto out;
7119                 }
7120                 /*
7121                  * when btrfs_get_extent can't find anything it
7122                  * returns one huge hole
7123                  *
7124                  * make sure what it found really fits our range, and
7125                  * adjust to make sure it is based on the start from
7126                  * the caller
7127                  */
7128                 if (hole_em) {
7129                         u64 calc_end = extent_map_end(hole_em);
7130
7131                         if (calc_end <= start || (hole_em->start > end)) {
7132                                 free_extent_map(hole_em);
7133                                 hole_em = NULL;
7134                         } else {
7135                                 hole_start = max(hole_em->start, start);
7136                                 hole_len = calc_end - hole_start;
7137                         }
7138                 }
7139                 em->bdev = NULL;
7140                 if (hole_em && range_start > hole_start) {
7141                         /* our hole starts before our delalloc, so we
7142                          * have to return just the parts of the hole
7143                          * that go until  the delalloc starts
7144                          */
7145                         em->len = min(hole_len,
7146                                       range_start - hole_start);
7147                         em->start = hole_start;
7148                         em->orig_start = hole_start;
7149                         /*
7150                          * don't adjust block start at all,
7151                          * it is fixed at EXTENT_MAP_HOLE
7152                          */
7153                         em->block_start = hole_em->block_start;
7154                         em->block_len = hole_len;
7155                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7156                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7157                 } else {
7158                         em->start = range_start;
7159                         em->len = found;
7160                         em->orig_start = range_start;
7161                         em->block_start = EXTENT_MAP_DELALLOC;
7162                         em->block_len = found;
7163                 }
7164         } else if (hole_em) {
7165                 return hole_em;
7166         }
7167 out:
7168
7169         free_extent_map(hole_em);
7170         if (err) {
7171                 free_extent_map(em);
7172                 return ERR_PTR(err);
7173         }
7174         return em;
7175 }
7176
7177 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7178                                                   const u64 start,
7179                                                   const u64 len,
7180                                                   const u64 orig_start,
7181                                                   const u64 block_start,
7182                                                   const u64 block_len,
7183                                                   const u64 orig_block_len,
7184                                                   const u64 ram_bytes,
7185                                                   const int type)
7186 {
7187         struct extent_map *em = NULL;
7188         int ret;
7189
7190         down_read(&BTRFS_I(inode)->dio_sem);
7191         if (type != BTRFS_ORDERED_NOCOW) {
7192                 em = create_pinned_em(inode, start, len, orig_start,
7193                                       block_start, block_len, orig_block_len,
7194                                       ram_bytes, type);
7195                 if (IS_ERR(em))
7196                         goto out;
7197         }
7198         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7199                                            len, block_len, type);
7200         if (ret) {
7201                 if (em) {
7202                         free_extent_map(em);
7203                         btrfs_drop_extent_cache(inode, start,
7204                                                 start + len - 1, 0);
7205                 }
7206                 em = ERR_PTR(ret);
7207         }
7208  out:
7209         up_read(&BTRFS_I(inode)->dio_sem);
7210
7211         return em;
7212 }
7213
7214 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7215                                                   u64 start, u64 len)
7216 {
7217         struct btrfs_root *root = BTRFS_I(inode)->root;
7218         struct extent_map *em;
7219         struct btrfs_key ins;
7220         u64 alloc_hint;
7221         int ret;
7222
7223         alloc_hint = get_extent_allocation_hint(inode, start, len);
7224         ret = btrfs_reserve_extent(root, len, root->sectorsize, 0,
7225                                    alloc_hint, &ins, 1, 1);
7226         if (ret)
7227                 return ERR_PTR(ret);
7228
7229         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7230                                      ins.objectid, ins.offset, ins.offset,
7231                                      ins.offset, 0);
7232         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
7233         if (IS_ERR(em))
7234                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7235
7236         return em;
7237 }
7238
7239 /*
7240  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7241  * block must be cow'd
7242  */
7243 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7244                               u64 *orig_start, u64 *orig_block_len,
7245                               u64 *ram_bytes)
7246 {
7247         struct btrfs_trans_handle *trans;
7248         struct btrfs_path *path;
7249         int ret;
7250         struct extent_buffer *leaf;
7251         struct btrfs_root *root = BTRFS_I(inode)->root;
7252         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7253         struct btrfs_file_extent_item *fi;
7254         struct btrfs_key key;
7255         u64 disk_bytenr;
7256         u64 backref_offset;
7257         u64 extent_end;
7258         u64 num_bytes;
7259         int slot;
7260         int found_type;
7261         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7262
7263         path = btrfs_alloc_path();
7264         if (!path)
7265                 return -ENOMEM;
7266
7267         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode),
7268                                        offset, 0);
7269         if (ret < 0)
7270                 goto out;
7271
7272         slot = path->slots[0];
7273         if (ret == 1) {
7274                 if (slot == 0) {
7275                         /* can't find the item, must cow */
7276                         ret = 0;
7277                         goto out;
7278                 }
7279                 slot--;
7280         }
7281         ret = 0;
7282         leaf = path->nodes[0];
7283         btrfs_item_key_to_cpu(leaf, &key, slot);
7284         if (key.objectid != btrfs_ino(inode) ||
7285             key.type != BTRFS_EXTENT_DATA_KEY) {
7286                 /* not our file or wrong item type, must cow */
7287                 goto out;
7288         }
7289
7290         if (key.offset > offset) {
7291                 /* Wrong offset, must cow */
7292                 goto out;
7293         }
7294
7295         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7296         found_type = btrfs_file_extent_type(leaf, fi);
7297         if (found_type != BTRFS_FILE_EXTENT_REG &&
7298             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7299                 /* not a regular extent, must cow */
7300                 goto out;
7301         }
7302
7303         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7304                 goto out;
7305
7306         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7307         if (extent_end <= offset)
7308                 goto out;
7309
7310         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7311         if (disk_bytenr == 0)
7312                 goto out;
7313
7314         if (btrfs_file_extent_compression(leaf, fi) ||
7315             btrfs_file_extent_encryption(leaf, fi) ||
7316             btrfs_file_extent_other_encoding(leaf, fi))
7317                 goto out;
7318
7319         backref_offset = btrfs_file_extent_offset(leaf, fi);
7320
7321         if (orig_start) {
7322                 *orig_start = key.offset - backref_offset;
7323                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7324                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7325         }
7326
7327         if (btrfs_extent_readonly(root, disk_bytenr))
7328                 goto out;
7329
7330         num_bytes = min(offset + *len, extent_end) - offset;
7331         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7332                 u64 range_end;
7333
7334                 range_end = round_up(offset + num_bytes, root->sectorsize) - 1;
7335                 ret = test_range_bit(io_tree, offset, range_end,
7336                                      EXTENT_DELALLOC, 0, NULL);
7337                 if (ret) {
7338                         ret = -EAGAIN;
7339                         goto out;
7340                 }
7341         }
7342
7343         btrfs_release_path(path);
7344
7345         /*
7346          * look for other files referencing this extent, if we
7347          * find any we must cow
7348          */
7349         trans = btrfs_join_transaction(root);
7350         if (IS_ERR(trans)) {
7351                 ret = 0;
7352                 goto out;
7353         }
7354
7355         ret = btrfs_cross_ref_exist(trans, root, btrfs_ino(inode),
7356                                     key.offset - backref_offset, disk_bytenr);
7357         btrfs_end_transaction(trans, root);
7358         if (ret) {
7359                 ret = 0;
7360                 goto out;
7361         }
7362
7363         /*
7364          * adjust disk_bytenr and num_bytes to cover just the bytes
7365          * in this extent we are about to write.  If there
7366          * are any csums in that range we have to cow in order
7367          * to keep the csums correct
7368          */
7369         disk_bytenr += backref_offset;
7370         disk_bytenr += offset - key.offset;
7371         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
7372                                 goto out;
7373         /*
7374          * all of the above have passed, it is safe to overwrite this extent
7375          * without cow
7376          */
7377         *len = num_bytes;
7378         ret = 1;
7379 out:
7380         btrfs_free_path(path);
7381         return ret;
7382 }
7383
7384 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7385 {
7386         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7387         int found = false;
7388         void **pagep = NULL;
7389         struct page *page = NULL;
7390         int start_idx;
7391         int end_idx;
7392
7393         start_idx = start >> PAGE_SHIFT;
7394
7395         /*
7396          * end is the last byte in the last page.  end == start is legal
7397          */
7398         end_idx = end >> PAGE_SHIFT;
7399
7400         rcu_read_lock();
7401
7402         /* Most of the code in this while loop is lifted from
7403          * find_get_page.  It's been modified to begin searching from a
7404          * page and return just the first page found in that range.  If the
7405          * found idx is less than or equal to the end idx then we know that
7406          * a page exists.  If no pages are found or if those pages are
7407          * outside of the range then we're fine (yay!) */
7408         while (page == NULL &&
7409                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7410                 page = radix_tree_deref_slot(pagep);
7411                 if (unlikely(!page))
7412                         break;
7413
7414                 if (radix_tree_exception(page)) {
7415                         if (radix_tree_deref_retry(page)) {
7416                                 page = NULL;
7417                                 continue;
7418                         }
7419                         /*
7420                          * Otherwise, shmem/tmpfs must be storing a swap entry
7421                          * here as an exceptional entry: so return it without
7422                          * attempting to raise page count.
7423                          */
7424                         page = NULL;
7425                         break; /* TODO: Is this relevant for this use case? */
7426                 }
7427
7428                 if (!page_cache_get_speculative(page)) {
7429                         page = NULL;
7430                         continue;
7431                 }
7432
7433                 /*
7434                  * Has the page moved?
7435                  * This is part of the lockless pagecache protocol. See
7436                  * include/linux/pagemap.h for details.
7437                  */
7438                 if (unlikely(page != *pagep)) {
7439                         put_page(page);
7440                         page = NULL;
7441                 }
7442         }
7443
7444         if (page) {
7445                 if (page->index <= end_idx)
7446                         found = true;
7447                 put_page(page);
7448         }
7449
7450         rcu_read_unlock();
7451         return found;
7452 }
7453
7454 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7455                               struct extent_state **cached_state, int writing)
7456 {
7457         struct btrfs_ordered_extent *ordered;
7458         int ret = 0;
7459
7460         while (1) {
7461                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7462                                  cached_state);
7463                 /*
7464                  * We're concerned with the entire range that we're going to be
7465                  * doing DIO to, so we need to make sure there's no ordered
7466                  * extents in this range.
7467                  */
7468                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
7469                                                      lockend - lockstart + 1);
7470
7471                 /*
7472                  * We need to make sure there are no buffered pages in this
7473                  * range either, we could have raced between the invalidate in
7474                  * generic_file_direct_write and locking the extent.  The
7475                  * invalidate needs to happen so that reads after a write do not
7476                  * get stale data.
7477                  */
7478                 if (!ordered &&
7479                     (!writing ||
7480                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7481                         break;
7482
7483                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7484                                      cached_state, GFP_NOFS);
7485
7486                 if (ordered) {
7487                         /*
7488                          * If we are doing a DIO read and the ordered extent we
7489                          * found is for a buffered write, we can not wait for it
7490                          * to complete and retry, because if we do so we can
7491                          * deadlock with concurrent buffered writes on page
7492                          * locks. This happens only if our DIO read covers more
7493                          * than one extent map, if at this point has already
7494                          * created an ordered extent for a previous extent map
7495                          * and locked its range in the inode's io tree, and a
7496                          * concurrent write against that previous extent map's
7497                          * range and this range started (we unlock the ranges
7498                          * in the io tree only when the bios complete and
7499                          * buffered writes always lock pages before attempting
7500                          * to lock range in the io tree).
7501                          */
7502                         if (writing ||
7503                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7504                                 btrfs_start_ordered_extent(inode, ordered, 1);
7505                         else
7506                                 ret = -ENOTBLK;
7507                         btrfs_put_ordered_extent(ordered);
7508                 } else {
7509                         /*
7510                          * We could trigger writeback for this range (and wait
7511                          * for it to complete) and then invalidate the pages for
7512                          * this range (through invalidate_inode_pages2_range()),
7513                          * but that can lead us to a deadlock with a concurrent
7514                          * call to readpages() (a buffered read or a defrag call
7515                          * triggered a readahead) on a page lock due to an
7516                          * ordered dio extent we created before but did not have
7517                          * yet a corresponding bio submitted (whence it can not
7518                          * complete), which makes readpages() wait for that
7519                          * ordered extent to complete while holding a lock on
7520                          * that page.
7521                          */
7522                         ret = -ENOTBLK;
7523                 }
7524
7525                 if (ret)
7526                         break;
7527
7528                 cond_resched();
7529         }
7530
7531         return ret;
7532 }
7533
7534 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
7535                                            u64 len, u64 orig_start,
7536                                            u64 block_start, u64 block_len,
7537                                            u64 orig_block_len, u64 ram_bytes,
7538                                            int type)
7539 {
7540         struct extent_map_tree *em_tree;
7541         struct extent_map *em;
7542         struct btrfs_root *root = BTRFS_I(inode)->root;
7543         int ret;
7544
7545         em_tree = &BTRFS_I(inode)->extent_tree;
7546         em = alloc_extent_map();
7547         if (!em)
7548                 return ERR_PTR(-ENOMEM);
7549
7550         em->start = start;
7551         em->orig_start = orig_start;
7552         em->mod_start = start;
7553         em->mod_len = len;
7554         em->len = len;
7555         em->block_len = block_len;
7556         em->block_start = block_start;
7557         em->bdev = root->fs_info->fs_devices->latest_bdev;
7558         em->orig_block_len = orig_block_len;
7559         em->ram_bytes = ram_bytes;
7560         em->generation = -1;
7561         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7562         if (type == BTRFS_ORDERED_PREALLOC)
7563                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7564
7565         do {
7566                 btrfs_drop_extent_cache(inode, em->start,
7567                                 em->start + em->len - 1, 0);
7568                 write_lock(&em_tree->lock);
7569                 ret = add_extent_mapping(em_tree, em, 1);
7570                 write_unlock(&em_tree->lock);
7571         } while (ret == -EEXIST);
7572
7573         if (ret) {
7574                 free_extent_map(em);
7575                 return ERR_PTR(ret);
7576         }
7577
7578         return em;
7579 }
7580
7581 static void adjust_dio_outstanding_extents(struct inode *inode,
7582                                            struct btrfs_dio_data *dio_data,
7583                                            const u64 len)
7584 {
7585         unsigned num_extents;
7586
7587         num_extents = (unsigned) div64_u64(len + BTRFS_MAX_EXTENT_SIZE - 1,
7588                                            BTRFS_MAX_EXTENT_SIZE);
7589         /*
7590          * If we have an outstanding_extents count still set then we're
7591          * within our reservation, otherwise we need to adjust our inode
7592          * counter appropriately.
7593          */
7594         if (dio_data->outstanding_extents) {
7595                 dio_data->outstanding_extents -= num_extents;
7596         } else {
7597                 spin_lock(&BTRFS_I(inode)->lock);
7598                 BTRFS_I(inode)->outstanding_extents += num_extents;
7599                 spin_unlock(&BTRFS_I(inode)->lock);
7600         }
7601 }
7602
7603 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7604                                    struct buffer_head *bh_result, int create)
7605 {
7606         struct extent_map *em;
7607         struct btrfs_root *root = BTRFS_I(inode)->root;
7608         struct extent_state *cached_state = NULL;
7609         struct btrfs_dio_data *dio_data = NULL;
7610         u64 start = iblock << inode->i_blkbits;
7611         u64 lockstart, lockend;
7612         u64 len = bh_result->b_size;
7613         int unlock_bits = EXTENT_LOCKED;
7614         int ret = 0;
7615
7616         if (create)
7617                 unlock_bits |= EXTENT_DIRTY;
7618         else
7619                 len = min_t(u64, len, root->sectorsize);
7620
7621         lockstart = start;
7622         lockend = start + len - 1;
7623
7624         if (current->journal_info) {
7625                 /*
7626                  * Need to pull our outstanding extents and set journal_info to NULL so
7627                  * that anything that needs to check if there's a transaction doesn't get
7628                  * confused.
7629                  */
7630                 dio_data = current->journal_info;
7631                 current->journal_info = NULL;
7632         }
7633
7634         /*
7635          * If this errors out it's because we couldn't invalidate pagecache for
7636          * this range and we need to fallback to buffered.
7637          */
7638         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7639                                create)) {
7640                 ret = -ENOTBLK;
7641                 goto err;
7642         }
7643
7644         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7645         if (IS_ERR(em)) {
7646                 ret = PTR_ERR(em);
7647                 goto unlock_err;
7648         }
7649
7650         /*
7651          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7652          * io.  INLINE is special, and we could probably kludge it in here, but
7653          * it's still buffered so for safety lets just fall back to the generic
7654          * buffered path.
7655          *
7656          * For COMPRESSED we _have_ to read the entire extent in so we can
7657          * decompress it, so there will be buffering required no matter what we
7658          * do, so go ahead and fallback to buffered.
7659          *
7660          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7661          * to buffered IO.  Don't blame me, this is the price we pay for using
7662          * the generic code.
7663          */
7664         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7665             em->block_start == EXTENT_MAP_INLINE) {
7666                 free_extent_map(em);
7667                 ret = -ENOTBLK;
7668                 goto unlock_err;
7669         }
7670
7671         /* Just a good old fashioned hole, return */
7672         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7673                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7674                 free_extent_map(em);
7675                 goto unlock_err;
7676         }
7677
7678         /*
7679          * We don't allocate a new extent in the following cases
7680          *
7681          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7682          * existing extent.
7683          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7684          * just use the extent.
7685          *
7686          */
7687         if (!create) {
7688                 len = min(len, em->len - (start - em->start));
7689                 lockstart = start + len;
7690                 goto unlock;
7691         }
7692
7693         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7694             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7695              em->block_start != EXTENT_MAP_HOLE)) {
7696                 int type;
7697                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7698
7699                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7700                         type = BTRFS_ORDERED_PREALLOC;
7701                 else
7702                         type = BTRFS_ORDERED_NOCOW;
7703                 len = min(len, em->len - (start - em->start));
7704                 block_start = em->block_start + (start - em->start);
7705
7706                 if (can_nocow_extent(inode, start, &len, &orig_start,
7707                                      &orig_block_len, &ram_bytes) == 1 &&
7708                     btrfs_inc_nocow_writers(root->fs_info, block_start)) {
7709                         struct extent_map *em2;
7710
7711                         em2 = btrfs_create_dio_extent(inode, start, len,
7712                                                       orig_start, block_start,
7713                                                       len, orig_block_len,
7714                                                       ram_bytes, type);
7715                         btrfs_dec_nocow_writers(root->fs_info, block_start);
7716                         if (type == BTRFS_ORDERED_PREALLOC) {
7717                                 free_extent_map(em);
7718                                 em = em2;
7719                         }
7720                         if (em2 && IS_ERR(em2)) {
7721                                 ret = PTR_ERR(em2);
7722                                 goto unlock_err;
7723                         }
7724                         goto unlock;
7725                 }
7726         }
7727
7728         /*
7729          * this will cow the extent, reset the len in case we changed
7730          * it above
7731          */
7732         len = bh_result->b_size;
7733         free_extent_map(em);
7734         em = btrfs_new_extent_direct(inode, start, len);
7735         if (IS_ERR(em)) {
7736                 ret = PTR_ERR(em);
7737                 goto unlock_err;
7738         }
7739         len = min(len, em->len - (start - em->start));
7740 unlock:
7741         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7742                 inode->i_blkbits;
7743         bh_result->b_size = len;
7744         bh_result->b_bdev = em->bdev;
7745         set_buffer_mapped(bh_result);
7746         if (create) {
7747                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7748                         set_buffer_new(bh_result);
7749
7750                 /*
7751                  * Need to update the i_size under the extent lock so buffered
7752                  * readers will get the updated i_size when we unlock.
7753                  */
7754                 if (start + len > i_size_read(inode))
7755                         i_size_write(inode, start + len);
7756
7757                 adjust_dio_outstanding_extents(inode, dio_data, len);
7758                 btrfs_free_reserved_data_space(inode, start, len);
7759                 WARN_ON(dio_data->reserve < len);
7760                 dio_data->reserve -= len;
7761                 dio_data->unsubmitted_oe_range_end = start + len;
7762                 current->journal_info = dio_data;
7763         }
7764
7765         /*
7766          * In the case of write we need to clear and unlock the entire range,
7767          * in the case of read we need to unlock only the end area that we
7768          * aren't using if there is any left over space.
7769          */
7770         if (lockstart < lockend) {
7771                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7772                                  lockend, unlock_bits, 1, 0,
7773                                  &cached_state, GFP_NOFS);
7774         } else {
7775                 free_extent_state(cached_state);
7776         }
7777
7778         free_extent_map(em);
7779
7780         return 0;
7781
7782 unlock_err:
7783         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7784                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7785 err:
7786         if (dio_data)
7787                 current->journal_info = dio_data;
7788         /*
7789          * Compensate the delalloc release we do in btrfs_direct_IO() when we
7790          * write less data then expected, so that we don't underflow our inode's
7791          * outstanding extents counter.
7792          */
7793         if (create && dio_data)
7794                 adjust_dio_outstanding_extents(inode, dio_data, len);
7795
7796         return ret;
7797 }
7798
7799 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7800                                         int rw, int mirror_num)
7801 {
7802         struct btrfs_root *root = BTRFS_I(inode)->root;
7803         int ret;
7804
7805         BUG_ON(rw & REQ_WRITE);
7806
7807         bio_get(bio);
7808
7809         ret = btrfs_bio_wq_end_io(root->fs_info, bio,
7810                                   BTRFS_WQ_ENDIO_DIO_REPAIR);
7811         if (ret)
7812                 goto err;
7813
7814         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
7815 err:
7816         bio_put(bio);
7817         return ret;
7818 }
7819
7820 static int btrfs_check_dio_repairable(struct inode *inode,
7821                                       struct bio *failed_bio,
7822                                       struct io_failure_record *failrec,
7823                                       int failed_mirror)
7824 {
7825         int num_copies;
7826
7827         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
7828                                       failrec->logical, failrec->len);
7829         if (num_copies == 1) {
7830                 /*
7831                  * we only have a single copy of the data, so don't bother with
7832                  * all the retry and error correction code that follows. no
7833                  * matter what the error is, it is very likely to persist.
7834                  */
7835                 pr_debug("Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
7836                          num_copies, failrec->this_mirror, failed_mirror);
7837                 return 0;
7838         }
7839
7840         failrec->failed_mirror = failed_mirror;
7841         failrec->this_mirror++;
7842         if (failrec->this_mirror == failed_mirror)
7843                 failrec->this_mirror++;
7844
7845         if (failrec->this_mirror > num_copies) {
7846                 pr_debug("Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
7847                          num_copies, failrec->this_mirror, failed_mirror);
7848                 return 0;
7849         }
7850
7851         return 1;
7852 }
7853
7854 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
7855                         struct page *page, unsigned int pgoff,
7856                         u64 start, u64 end, int failed_mirror,
7857                         bio_end_io_t *repair_endio, void *repair_arg)
7858 {
7859         struct io_failure_record *failrec;
7860         struct bio *bio;
7861         int isector;
7862         int read_mode;
7863         int ret;
7864
7865         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
7866
7867         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7868         if (ret)
7869                 return ret;
7870
7871         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7872                                          failed_mirror);
7873         if (!ret) {
7874                 free_io_failure(inode, failrec);
7875                 return -EIO;
7876         }
7877
7878         if ((failed_bio->bi_vcnt > 1)
7879                 || (failed_bio->bi_io_vec->bv_len
7880                         > BTRFS_I(inode)->root->sectorsize))
7881                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
7882         else
7883                 read_mode = READ_SYNC;
7884
7885         isector = start - btrfs_io_bio(failed_bio)->logical;
7886         isector >>= inode->i_sb->s_blocksize_bits;
7887         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7888                                 pgoff, isector, repair_endio, repair_arg);
7889         if (!bio) {
7890                 free_io_failure(inode, failrec);
7891                 return -EIO;
7892         }
7893
7894         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7895                     "Repair DIO Read Error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d\n",
7896                     read_mode, failrec->this_mirror, failrec->in_validation);
7897
7898         ret = submit_dio_repair_bio(inode, bio, read_mode,
7899                                     failrec->this_mirror);
7900         if (ret) {
7901                 free_io_failure(inode, failrec);
7902                 bio_put(bio);
7903         }
7904
7905         return ret;
7906 }
7907
7908 struct btrfs_retry_complete {
7909         struct completion done;
7910         struct inode *inode;
7911         u64 start;
7912         int uptodate;
7913 };
7914
7915 static void btrfs_retry_endio_nocsum(struct bio *bio)
7916 {
7917         struct btrfs_retry_complete *done = bio->bi_private;
7918         struct inode *inode;
7919         struct bio_vec *bvec;
7920         int i;
7921
7922         if (bio->bi_error)
7923                 goto end;
7924
7925         ASSERT(bio->bi_vcnt == 1);
7926         inode = bio->bi_io_vec->bv_page->mapping->host;
7927         ASSERT(bio->bi_io_vec->bv_len == BTRFS_I(inode)->root->sectorsize);
7928
7929         done->uptodate = 1;
7930         bio_for_each_segment_all(bvec, bio, i)
7931                 clean_io_failure(done->inode, done->start, bvec->bv_page, 0);
7932 end:
7933         complete(&done->done);
7934         bio_put(bio);
7935 }
7936
7937 static int __btrfs_correct_data_nocsum(struct inode *inode,
7938                                        struct btrfs_io_bio *io_bio)
7939 {
7940         struct btrfs_fs_info *fs_info;
7941         struct bio_vec *bvec;
7942         struct btrfs_retry_complete done;
7943         u64 start;
7944         unsigned int pgoff;
7945         u32 sectorsize;
7946         int nr_sectors;
7947         int i;
7948         int ret;
7949
7950         fs_info = BTRFS_I(inode)->root->fs_info;
7951         sectorsize = BTRFS_I(inode)->root->sectorsize;
7952
7953         start = io_bio->logical;
7954         done.inode = inode;
7955
7956         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7957                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
7958                 pgoff = bvec->bv_offset;
7959
7960 next_block_or_try_again:
7961                 done.uptodate = 0;
7962                 done.start = start;
7963                 init_completion(&done.done);
7964
7965                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
7966                                 pgoff, start, start + sectorsize - 1,
7967                                 io_bio->mirror_num,
7968                                 btrfs_retry_endio_nocsum, &done);
7969                 if (ret)
7970                         return ret;
7971
7972                 wait_for_completion(&done.done);
7973
7974                 if (!done.uptodate) {
7975                         /* We might have another mirror, so try again */
7976                         goto next_block_or_try_again;
7977                 }
7978
7979                 start += sectorsize;
7980
7981                 if (nr_sectors--) {
7982                         pgoff += sectorsize;
7983                         goto next_block_or_try_again;
7984                 }
7985         }
7986
7987         return 0;
7988 }
7989
7990 static void btrfs_retry_endio(struct bio *bio)
7991 {
7992         struct btrfs_retry_complete *done = bio->bi_private;
7993         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7994         struct inode *inode;
7995         struct bio_vec *bvec;
7996         u64 start;
7997         int uptodate;
7998         int ret;
7999         int i;
8000
8001         if (bio->bi_error)
8002                 goto end;
8003
8004         uptodate = 1;
8005
8006         start = done->start;
8007
8008         ASSERT(bio->bi_vcnt == 1);
8009         inode = bio->bi_io_vec->bv_page->mapping->host;
8010         ASSERT(bio->bi_io_vec->bv_len == BTRFS_I(inode)->root->sectorsize);
8011
8012         bio_for_each_segment_all(bvec, bio, i) {
8013                 ret = __readpage_endio_check(done->inode, io_bio, i,
8014                                         bvec->bv_page, bvec->bv_offset,
8015                                         done->start, bvec->bv_len);
8016                 if (!ret)
8017                         clean_io_failure(done->inode, done->start,
8018                                         bvec->bv_page, bvec->bv_offset);
8019                 else
8020                         uptodate = 0;
8021         }
8022
8023         done->uptodate = uptodate;
8024 end:
8025         complete(&done->done);
8026         bio_put(bio);
8027 }
8028
8029 static int __btrfs_subio_endio_read(struct inode *inode,
8030                                     struct btrfs_io_bio *io_bio, int err)
8031 {
8032         struct btrfs_fs_info *fs_info;
8033         struct bio_vec *bvec;
8034         struct btrfs_retry_complete done;
8035         u64 start;
8036         u64 offset = 0;
8037         u32 sectorsize;
8038         int nr_sectors;
8039         unsigned int pgoff;
8040         int csum_pos;
8041         int i;
8042         int ret;
8043
8044         fs_info = BTRFS_I(inode)->root->fs_info;
8045         sectorsize = BTRFS_I(inode)->root->sectorsize;
8046
8047         err = 0;
8048         start = io_bio->logical;
8049         done.inode = inode;
8050
8051         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
8052                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
8053
8054                 pgoff = bvec->bv_offset;
8055 next_block:
8056                 csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8057                 ret = __readpage_endio_check(inode, io_bio, csum_pos,
8058                                         bvec->bv_page, pgoff, start,
8059                                         sectorsize);
8060                 if (likely(!ret))
8061                         goto next;
8062 try_again:
8063                 done.uptodate = 0;
8064                 done.start = start;
8065                 init_completion(&done.done);
8066
8067                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
8068                                 pgoff, start, start + sectorsize - 1,
8069                                 io_bio->mirror_num,
8070                                 btrfs_retry_endio, &done);
8071                 if (ret) {
8072                         err = ret;
8073                         goto next;
8074                 }
8075
8076                 wait_for_completion(&done.done);
8077
8078                 if (!done.uptodate) {
8079                         /* We might have another mirror, so try again */
8080                         goto try_again;
8081                 }
8082 next:
8083                 offset += sectorsize;
8084                 start += sectorsize;
8085
8086                 ASSERT(nr_sectors);
8087
8088                 if (--nr_sectors) {
8089                         pgoff += sectorsize;
8090                         goto next_block;
8091                 }
8092         }
8093
8094         return err;
8095 }
8096
8097 static int btrfs_subio_endio_read(struct inode *inode,
8098                                   struct btrfs_io_bio *io_bio, int err)
8099 {
8100         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8101
8102         if (skip_csum) {
8103                 if (unlikely(err))
8104                         return __btrfs_correct_data_nocsum(inode, io_bio);
8105                 else
8106                         return 0;
8107         } else {
8108                 return __btrfs_subio_endio_read(inode, io_bio, err);
8109         }
8110 }
8111
8112 static void btrfs_endio_direct_read(struct bio *bio)
8113 {
8114         struct btrfs_dio_private *dip = bio->bi_private;
8115         struct inode *inode = dip->inode;
8116         struct bio *dio_bio;
8117         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8118         int err = bio->bi_error;
8119
8120         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8121                 err = btrfs_subio_endio_read(inode, io_bio, err);
8122
8123         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8124                       dip->logical_offset + dip->bytes - 1);
8125         dio_bio = dip->dio_bio;
8126
8127         kfree(dip);
8128
8129         dio_bio->bi_error = bio->bi_error;
8130         dio_end_io(dio_bio, bio->bi_error);
8131
8132         if (io_bio->end_io)
8133                 io_bio->end_io(io_bio, err);
8134         bio_put(bio);
8135 }
8136
8137 static void btrfs_endio_direct_write_update_ordered(struct inode *inode,
8138                                                     const u64 offset,
8139                                                     const u64 bytes,
8140                                                     const int uptodate)
8141 {
8142         struct btrfs_root *root = BTRFS_I(inode)->root;
8143         struct btrfs_ordered_extent *ordered = NULL;
8144         u64 ordered_offset = offset;
8145         u64 ordered_bytes = bytes;
8146         int ret;
8147
8148 again:
8149         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8150                                                    &ordered_offset,
8151                                                    ordered_bytes,
8152                                                    uptodate);
8153         if (!ret)
8154                 goto out_test;
8155
8156         btrfs_init_work(&ordered->work, btrfs_endio_write_helper,
8157                         finish_ordered_fn, NULL, NULL);
8158         btrfs_queue_work(root->fs_info->endio_write_workers,
8159                          &ordered->work);
8160 out_test:
8161         /*
8162          * our bio might span multiple ordered extents.  If we haven't
8163          * completed the accounting for the whole dio, go back and try again
8164          */
8165         if (ordered_offset < offset + bytes) {
8166                 ordered_bytes = offset + bytes - ordered_offset;
8167                 ordered = NULL;
8168                 goto again;
8169         }
8170 }
8171
8172 static void btrfs_endio_direct_write(struct bio *bio)
8173 {
8174         struct btrfs_dio_private *dip = bio->bi_private;
8175         struct bio *dio_bio = dip->dio_bio;
8176
8177         btrfs_endio_direct_write_update_ordered(dip->inode,
8178                                                 dip->logical_offset,
8179                                                 dip->bytes,
8180                                                 !bio->bi_error);
8181
8182         kfree(dip);
8183
8184         dio_bio->bi_error = bio->bi_error;
8185         dio_end_io(dio_bio, bio->bi_error);
8186         bio_put(bio);
8187 }
8188
8189 static int __btrfs_submit_bio_start_direct_io(struct inode *inode, int rw,
8190                                     struct bio *bio, int mirror_num,
8191                                     unsigned long bio_flags, u64 offset)
8192 {
8193         int ret;
8194         struct btrfs_root *root = BTRFS_I(inode)->root;
8195         ret = btrfs_csum_one_bio(root, inode, bio, offset, 1);
8196         BUG_ON(ret); /* -ENOMEM */
8197         return 0;
8198 }
8199
8200 static void btrfs_end_dio_bio(struct bio *bio)
8201 {
8202         struct btrfs_dio_private *dip = bio->bi_private;
8203         int err = bio->bi_error;
8204
8205         if (err)
8206                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8207                            "direct IO failed ino %llu rw %lu sector %#Lx len %u err no %d",
8208                            btrfs_ino(dip->inode), bio->bi_rw,
8209                            (unsigned long long)bio->bi_iter.bi_sector,
8210                            bio->bi_iter.bi_size, err);
8211
8212         if (dip->subio_endio)
8213                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8214
8215         if (err) {
8216                 dip->errors = 1;
8217
8218                 /*
8219                  * before atomic variable goto zero, we must make sure
8220                  * dip->errors is perceived to be set.
8221                  */
8222                 smp_mb__before_atomic();
8223         }
8224
8225         /* if there are more bios still pending for this dio, just exit */
8226         if (!atomic_dec_and_test(&dip->pending_bios))
8227                 goto out;
8228
8229         if (dip->errors) {
8230                 bio_io_error(dip->orig_bio);
8231         } else {
8232                 dip->dio_bio->bi_error = 0;
8233                 bio_endio(dip->orig_bio);
8234         }
8235 out:
8236         bio_put(bio);
8237 }
8238
8239 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
8240                                        u64 first_sector, gfp_t gfp_flags)
8241 {
8242         struct bio *bio;
8243         bio = btrfs_bio_alloc(bdev, first_sector, BIO_MAX_PAGES, gfp_flags);
8244         if (bio)
8245                 bio_associate_current(bio);
8246         return bio;
8247 }
8248
8249 static inline int btrfs_lookup_and_bind_dio_csum(struct btrfs_root *root,
8250                                                  struct inode *inode,
8251                                                  struct btrfs_dio_private *dip,
8252                                                  struct bio *bio,
8253                                                  u64 file_offset)
8254 {
8255         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8256         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8257         int ret;
8258
8259         /*
8260          * We load all the csum data we need when we submit
8261          * the first bio to reduce the csum tree search and
8262          * contention.
8263          */
8264         if (dip->logical_offset == file_offset) {
8265                 ret = btrfs_lookup_bio_sums_dio(root, inode, dip->orig_bio,
8266                                                 file_offset);
8267                 if (ret)
8268                         return ret;
8269         }
8270
8271         if (bio == dip->orig_bio)
8272                 return 0;
8273
8274         file_offset -= dip->logical_offset;
8275         file_offset >>= inode->i_sb->s_blocksize_bits;
8276         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8277
8278         return 0;
8279 }
8280
8281 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8282                                          int rw, u64 file_offset, int skip_sum,
8283                                          int async_submit)
8284 {
8285         struct btrfs_dio_private *dip = bio->bi_private;
8286         int write = rw & REQ_WRITE;
8287         struct btrfs_root *root = BTRFS_I(inode)->root;
8288         int ret;
8289
8290         if (async_submit)
8291                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8292
8293         bio_get(bio);
8294
8295         if (!write) {
8296                 ret = btrfs_bio_wq_end_io(root->fs_info, bio,
8297                                 BTRFS_WQ_ENDIO_DATA);
8298                 if (ret)
8299                         goto err;
8300         }
8301
8302         if (skip_sum)
8303                 goto map;
8304
8305         if (write && async_submit) {
8306                 ret = btrfs_wq_submit_bio(root->fs_info,
8307                                    inode, rw, bio, 0, 0,
8308                                    file_offset,
8309                                    __btrfs_submit_bio_start_direct_io,
8310                                    __btrfs_submit_bio_done);
8311                 goto err;
8312         } else if (write) {
8313                 /*
8314                  * If we aren't doing async submit, calculate the csum of the
8315                  * bio now.
8316                  */
8317                 ret = btrfs_csum_one_bio(root, inode, bio, file_offset, 1);
8318                 if (ret)
8319                         goto err;
8320         } else {
8321                 ret = btrfs_lookup_and_bind_dio_csum(root, inode, dip, bio,
8322                                                      file_offset);
8323                 if (ret)
8324                         goto err;
8325         }
8326 map:
8327         ret = btrfs_map_bio(root, rw, bio, 0, async_submit);
8328 err:
8329         bio_put(bio);
8330         return ret;
8331 }
8332
8333 static int btrfs_submit_direct_hook(int rw, struct btrfs_dio_private *dip,
8334                                     int skip_sum)
8335 {
8336         struct inode *inode = dip->inode;
8337         struct btrfs_root *root = BTRFS_I(inode)->root;
8338         struct bio *bio;
8339         struct bio *orig_bio = dip->orig_bio;
8340         struct bio_vec *bvec = orig_bio->bi_io_vec;
8341         u64 start_sector = orig_bio->bi_iter.bi_sector;
8342         u64 file_offset = dip->logical_offset;
8343         u64 submit_len = 0;
8344         u64 map_length;
8345         u32 blocksize = root->sectorsize;
8346         int async_submit = 0;
8347         int nr_sectors;
8348         int ret;
8349         int i;
8350
8351         map_length = orig_bio->bi_iter.bi_size;
8352         ret = btrfs_map_block(root->fs_info, rw, start_sector << 9,
8353                               &map_length, NULL, 0);
8354         if (ret)
8355                 return -EIO;
8356
8357         if (map_length >= orig_bio->bi_iter.bi_size) {
8358                 bio = orig_bio;
8359                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8360                 goto submit;
8361         }
8362
8363         /* async crcs make it difficult to collect full stripe writes. */
8364         if (btrfs_get_alloc_profile(root, 1) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8365                 async_submit = 0;
8366         else
8367                 async_submit = 1;
8368
8369         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
8370         if (!bio)
8371                 return -ENOMEM;
8372
8373         bio->bi_private = dip;
8374         bio->bi_end_io = btrfs_end_dio_bio;
8375         btrfs_io_bio(bio)->logical = file_offset;
8376         atomic_inc(&dip->pending_bios);
8377
8378         while (bvec <= (orig_bio->bi_io_vec + orig_bio->bi_vcnt - 1)) {
8379                 nr_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info, bvec->bv_len);
8380                 i = 0;
8381 next_block:
8382                 if (unlikely(map_length < submit_len + blocksize ||
8383                     bio_add_page(bio, bvec->bv_page, blocksize,
8384                             bvec->bv_offset + (i * blocksize)) < blocksize)) {
8385                         /*
8386                          * inc the count before we submit the bio so
8387                          * we know the end IO handler won't happen before
8388                          * we inc the count. Otherwise, the dip might get freed
8389                          * before we're done setting it up
8390                          */
8391                         atomic_inc(&dip->pending_bios);
8392                         ret = __btrfs_submit_dio_bio(bio, inode, rw,
8393                                                      file_offset, skip_sum,
8394                                                      async_submit);
8395                         if (ret) {
8396                                 bio_put(bio);
8397                                 atomic_dec(&dip->pending_bios);
8398                                 goto out_err;
8399                         }
8400
8401                         start_sector += submit_len >> 9;
8402                         file_offset += submit_len;
8403
8404                         submit_len = 0;
8405
8406                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
8407                                                   start_sector, GFP_NOFS);
8408                         if (!bio)
8409                                 goto out_err;
8410                         bio->bi_private = dip;
8411                         bio->bi_end_io = btrfs_end_dio_bio;
8412                         btrfs_io_bio(bio)->logical = file_offset;
8413
8414                         map_length = orig_bio->bi_iter.bi_size;
8415                         ret = btrfs_map_block(root->fs_info, rw,
8416                                               start_sector << 9,
8417                                               &map_length, NULL, 0);
8418                         if (ret) {
8419                                 bio_put(bio);
8420                                 goto out_err;
8421                         }
8422
8423                         goto next_block;
8424                 } else {
8425                         submit_len += blocksize;
8426                         if (--nr_sectors) {
8427                                 i++;
8428                                 goto next_block;
8429                         }
8430                         bvec++;
8431                 }
8432         }
8433
8434 submit:
8435         ret = __btrfs_submit_dio_bio(bio, inode, rw, file_offset, skip_sum,
8436                                      async_submit);
8437         if (!ret)
8438                 return 0;
8439
8440         bio_put(bio);
8441 out_err:
8442         dip->errors = 1;
8443         /*
8444          * before atomic variable goto zero, we must
8445          * make sure dip->errors is perceived to be set.
8446          */
8447         smp_mb__before_atomic();
8448         if (atomic_dec_and_test(&dip->pending_bios))
8449                 bio_io_error(dip->orig_bio);
8450
8451         /* bio_end_io() will handle error, so we needn't return it */
8452         return 0;
8453 }
8454
8455 static void btrfs_submit_direct(int rw, struct bio *dio_bio,
8456                                 struct inode *inode, loff_t file_offset)
8457 {
8458         struct btrfs_dio_private *dip = NULL;
8459         struct bio *io_bio = NULL;
8460         struct btrfs_io_bio *btrfs_bio;
8461         int skip_sum;
8462         int write = rw & REQ_WRITE;
8463         int ret = 0;
8464
8465         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8466
8467         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
8468         if (!io_bio) {
8469                 ret = -ENOMEM;
8470                 goto free_ordered;
8471         }
8472
8473         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8474         if (!dip) {
8475                 ret = -ENOMEM;
8476                 goto free_ordered;
8477         }
8478
8479         dip->private = dio_bio->bi_private;
8480         dip->inode = inode;
8481         dip->logical_offset = file_offset;
8482         dip->bytes = dio_bio->bi_iter.bi_size;
8483         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8484         io_bio->bi_private = dip;
8485         dip->orig_bio = io_bio;
8486         dip->dio_bio = dio_bio;
8487         atomic_set(&dip->pending_bios, 0);
8488         btrfs_bio = btrfs_io_bio(io_bio);
8489         btrfs_bio->logical = file_offset;
8490
8491         if (write) {
8492                 io_bio->bi_end_io = btrfs_endio_direct_write;
8493         } else {
8494                 io_bio->bi_end_io = btrfs_endio_direct_read;
8495                 dip->subio_endio = btrfs_subio_endio_read;
8496         }
8497
8498         /*
8499          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8500          * even if we fail to submit a bio, because in such case we do the
8501          * corresponding error handling below and it must not be done a second
8502          * time by btrfs_direct_IO().
8503          */
8504         if (write) {
8505                 struct btrfs_dio_data *dio_data = current->journal_info;
8506
8507                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8508                         dip->bytes;
8509                 dio_data->unsubmitted_oe_range_start =
8510                         dio_data->unsubmitted_oe_range_end;
8511         }
8512
8513         ret = btrfs_submit_direct_hook(rw, dip, skip_sum);
8514         if (!ret)
8515                 return;
8516
8517         if (btrfs_bio->end_io)
8518                 btrfs_bio->end_io(btrfs_bio, ret);
8519
8520 free_ordered:
8521         /*
8522          * If we arrived here it means either we failed to submit the dip
8523          * or we either failed to clone the dio_bio or failed to allocate the
8524          * dip. If we cloned the dio_bio and allocated the dip, we can just
8525          * call bio_endio against our io_bio so that we get proper resource
8526          * cleanup if we fail to submit the dip, otherwise, we must do the
8527          * same as btrfs_endio_direct_[write|read] because we can't call these
8528          * callbacks - they require an allocated dip and a clone of dio_bio.
8529          */
8530         if (io_bio && dip) {
8531                 io_bio->bi_error = -EIO;
8532                 bio_endio(io_bio);
8533                 /*
8534                  * The end io callbacks free our dip, do the final put on io_bio
8535                  * and all the cleanup and final put for dio_bio (through
8536                  * dio_end_io()).
8537                  */
8538                 dip = NULL;
8539                 io_bio = NULL;
8540         } else {
8541                 if (write)
8542                         btrfs_endio_direct_write_update_ordered(inode,
8543                                                 file_offset,
8544                                                 dio_bio->bi_iter.bi_size,
8545                                                 0);
8546                 else
8547                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8548                               file_offset + dio_bio->bi_iter.bi_size - 1);
8549
8550                 dio_bio->bi_error = -EIO;
8551                 /*
8552                  * Releases and cleans up our dio_bio, no need to bio_put()
8553                  * nor bio_endio()/bio_io_error() against dio_bio.
8554                  */
8555                 dio_end_io(dio_bio, ret);
8556         }
8557         if (io_bio)
8558                 bio_put(io_bio);
8559         kfree(dip);
8560 }
8561
8562 static ssize_t check_direct_IO(struct btrfs_root *root, struct kiocb *iocb,
8563                         const struct iov_iter *iter, loff_t offset)
8564 {
8565         int seg;
8566         int i;
8567         unsigned blocksize_mask = root->sectorsize - 1;
8568         ssize_t retval = -EINVAL;
8569
8570         if (offset & blocksize_mask)
8571                 goto out;
8572
8573         if (iov_iter_alignment(iter) & blocksize_mask)
8574                 goto out;
8575
8576         /* If this is a write we don't need to check anymore */
8577         if (iov_iter_rw(iter) == WRITE)
8578                 return 0;
8579         /*
8580          * Check to make sure we don't have duplicate iov_base's in this
8581          * iovec, if so return EINVAL, otherwise we'll get csum errors
8582          * when reading back.
8583          */
8584         for (seg = 0; seg < iter->nr_segs; seg++) {
8585                 for (i = seg + 1; i < iter->nr_segs; i++) {
8586                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8587                                 goto out;
8588                 }
8589         }
8590         retval = 0;
8591 out:
8592         return retval;
8593 }
8594
8595 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8596 {
8597         struct file *file = iocb->ki_filp;
8598         struct inode *inode = file->f_mapping->host;
8599         struct btrfs_root *root = BTRFS_I(inode)->root;
8600         struct btrfs_dio_data dio_data = { 0 };
8601         loff_t offset = iocb->ki_pos;
8602         size_t count = 0;
8603         int flags = 0;
8604         bool wakeup = true;
8605         bool relock = false;
8606         ssize_t ret;
8607
8608         if (check_direct_IO(BTRFS_I(inode)->root, iocb, iter, offset))
8609                 return 0;
8610
8611         inode_dio_begin(inode);
8612         smp_mb__after_atomic();
8613
8614         /*
8615          * The generic stuff only does filemap_write_and_wait_range, which
8616          * isn't enough if we've written compressed pages to this area, so
8617          * we need to flush the dirty pages again to make absolutely sure
8618          * that any outstanding dirty pages are on disk.
8619          */
8620         count = iov_iter_count(iter);
8621         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8622                      &BTRFS_I(inode)->runtime_flags))
8623                 filemap_fdatawrite_range(inode->i_mapping, offset,
8624                                          offset + count - 1);
8625
8626         if (iov_iter_rw(iter) == WRITE) {
8627                 /*
8628                  * If the write DIO is beyond the EOF, we need update
8629                  * the isize, but it is protected by i_mutex. So we can
8630                  * not unlock the i_mutex at this case.
8631                  */
8632                 if (offset + count <= inode->i_size) {
8633                         inode_unlock(inode);
8634                         relock = true;
8635                 }
8636                 ret = btrfs_delalloc_reserve_space(inode, offset, count);
8637                 if (ret)
8638                         goto out;
8639                 dio_data.outstanding_extents = div64_u64(count +
8640                                                 BTRFS_MAX_EXTENT_SIZE - 1,
8641                                                 BTRFS_MAX_EXTENT_SIZE);
8642
8643                 /*
8644                  * We need to know how many extents we reserved so that we can
8645                  * do the accounting properly if we go over the number we
8646                  * originally calculated.  Abuse current->journal_info for this.
8647                  */
8648                 dio_data.reserve = round_up(count, root->sectorsize);
8649                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8650                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8651                 current->journal_info = &dio_data;
8652         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8653                                      &BTRFS_I(inode)->runtime_flags)) {
8654                 inode_dio_end(inode);
8655                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8656                 wakeup = false;
8657         }
8658
8659         ret = __blockdev_direct_IO(iocb, inode,
8660                                    BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev,
8661                                    iter, btrfs_get_blocks_direct, NULL,
8662                                    btrfs_submit_direct, flags);
8663         if (iov_iter_rw(iter) == WRITE) {
8664                 current->journal_info = NULL;
8665                 if (ret < 0 && ret != -EIOCBQUEUED) {
8666                         if (dio_data.reserve)
8667                                 btrfs_delalloc_release_space(inode, offset,
8668                                                              dio_data.reserve);
8669                         /*
8670                          * On error we might have left some ordered extents
8671                          * without submitting corresponding bios for them, so
8672                          * cleanup them up to avoid other tasks getting them
8673                          * and waiting for them to complete forever.
8674                          */
8675                         if (dio_data.unsubmitted_oe_range_start <
8676                             dio_data.unsubmitted_oe_range_end)
8677                                 btrfs_endio_direct_write_update_ordered(inode,
8678                                         dio_data.unsubmitted_oe_range_start,
8679                                         dio_data.unsubmitted_oe_range_end -
8680                                         dio_data.unsubmitted_oe_range_start,
8681                                         0);
8682                 } else if (ret >= 0 && (size_t)ret < count)
8683                         btrfs_delalloc_release_space(inode, offset,
8684                                                      count - (size_t)ret);
8685         }
8686 out:
8687         if (wakeup)
8688                 inode_dio_end(inode);
8689         if (relock)
8690                 inode_lock(inode);
8691
8692         return ret;
8693 }
8694
8695 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8696
8697 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8698                 __u64 start, __u64 len)
8699 {
8700         int     ret;
8701
8702         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8703         if (ret)
8704                 return ret;
8705
8706         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8707 }
8708
8709 int btrfs_readpage(struct file *file, struct page *page)
8710 {
8711         struct extent_io_tree *tree;
8712         tree = &BTRFS_I(page->mapping->host)->io_tree;
8713         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8714 }
8715
8716 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8717 {
8718         struct extent_io_tree *tree;
8719         struct inode *inode = page->mapping->host;
8720         int ret;
8721
8722         if (current->flags & PF_MEMALLOC) {
8723                 redirty_page_for_writepage(wbc, page);
8724                 unlock_page(page);
8725                 return 0;
8726         }
8727
8728         /*
8729          * If we are under memory pressure we will call this directly from the
8730          * VM, we need to make sure we have the inode referenced for the ordered
8731          * extent.  If not just return like we didn't do anything.
8732          */
8733         if (!igrab(inode)) {
8734                 redirty_page_for_writepage(wbc, page);
8735                 return AOP_WRITEPAGE_ACTIVATE;
8736         }
8737         tree = &BTRFS_I(page->mapping->host)->io_tree;
8738         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8739         btrfs_add_delayed_iput(inode);
8740         return ret;
8741 }
8742
8743 static int btrfs_writepages(struct address_space *mapping,
8744                             struct writeback_control *wbc)
8745 {
8746         struct extent_io_tree *tree;
8747
8748         tree = &BTRFS_I(mapping->host)->io_tree;
8749         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8750 }
8751
8752 static int
8753 btrfs_readpages(struct file *file, struct address_space *mapping,
8754                 struct list_head *pages, unsigned nr_pages)
8755 {
8756         struct extent_io_tree *tree;
8757         tree = &BTRFS_I(mapping->host)->io_tree;
8758         return extent_readpages(tree, mapping, pages, nr_pages,
8759                                 btrfs_get_extent);
8760 }
8761 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8762 {
8763         struct extent_io_tree *tree;
8764         struct extent_map_tree *map;
8765         int ret;
8766
8767         tree = &BTRFS_I(page->mapping->host)->io_tree;
8768         map = &BTRFS_I(page->mapping->host)->extent_tree;
8769         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8770         if (ret == 1) {
8771                 ClearPagePrivate(page);
8772                 set_page_private(page, 0);
8773                 put_page(page);
8774         }
8775         return ret;
8776 }
8777
8778 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8779 {
8780         if (PageWriteback(page) || PageDirty(page))
8781                 return 0;
8782         return __btrfs_releasepage(page, gfp_flags & GFP_NOFS);
8783 }
8784
8785 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8786                                  unsigned int length)
8787 {
8788         struct inode *inode = page->mapping->host;
8789         struct extent_io_tree *tree;
8790         struct btrfs_ordered_extent *ordered;
8791         struct extent_state *cached_state = NULL;
8792         u64 page_start = page_offset(page);
8793         u64 page_end = page_start + PAGE_SIZE - 1;
8794         u64 start;
8795         u64 end;
8796         int inode_evicting = inode->i_state & I_FREEING;
8797
8798         /*
8799          * we have the page locked, so new writeback can't start,
8800          * and the dirty bit won't be cleared while we are here.
8801          *
8802          * Wait for IO on this page so that we can safely clear
8803          * the PagePrivate2 bit and do ordered accounting
8804          */
8805         wait_on_page_writeback(page);
8806
8807         tree = &BTRFS_I(inode)->io_tree;
8808         if (offset) {
8809                 btrfs_releasepage(page, GFP_NOFS);
8810                 return;
8811         }
8812
8813         if (!inode_evicting)
8814                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8815 again:
8816         start = page_start;
8817         ordered = btrfs_lookup_ordered_range(inode, start,
8818                                         page_end - start + 1);
8819         if (ordered) {
8820                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8821                 /*
8822                  * IO on this page will never be started, so we need
8823                  * to account for any ordered extents now
8824                  */
8825                 if (!inode_evicting)
8826                         clear_extent_bit(tree, start, end,
8827                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8828                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8829                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8830                                          GFP_NOFS);
8831                 /*
8832                  * whoever cleared the private bit is responsible
8833                  * for the finish_ordered_io
8834                  */
8835                 if (TestClearPagePrivate2(page)) {
8836                         struct btrfs_ordered_inode_tree *tree;
8837                         u64 new_len;
8838
8839                         tree = &BTRFS_I(inode)->ordered_tree;
8840
8841                         spin_lock_irq(&tree->lock);
8842                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8843                         new_len = start - ordered->file_offset;
8844                         if (new_len < ordered->truncated_len)
8845                                 ordered->truncated_len = new_len;
8846                         spin_unlock_irq(&tree->lock);
8847
8848                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8849                                                            start,
8850                                                            end - start + 1, 1))
8851                                 btrfs_finish_ordered_io(ordered);
8852                 }
8853                 btrfs_put_ordered_extent(ordered);
8854                 if (!inode_evicting) {
8855                         cached_state = NULL;
8856                         lock_extent_bits(tree, start, end,
8857                                          &cached_state);
8858                 }
8859
8860                 start = end + 1;
8861                 if (start < page_end)
8862                         goto again;
8863         }
8864
8865         /*
8866          * Qgroup reserved space handler
8867          * Page here will be either
8868          * 1) Already written to disk
8869          *    In this case, its reserved space is released from data rsv map
8870          *    and will be freed by delayed_ref handler finally.
8871          *    So even we call qgroup_free_data(), it won't decrease reserved
8872          *    space.
8873          * 2) Not written to disk
8874          *    This means the reserved space should be freed here.
8875          */
8876         btrfs_qgroup_free_data(inode, page_start, PAGE_SIZE);
8877         if (!inode_evicting) {
8878                 clear_extent_bit(tree, page_start, page_end,
8879                                  EXTENT_LOCKED | EXTENT_DIRTY |
8880                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8881                                  EXTENT_DEFRAG, 1, 1,
8882                                  &cached_state, GFP_NOFS);
8883
8884                 __btrfs_releasepage(page, GFP_NOFS);
8885         }
8886
8887         ClearPageChecked(page);
8888         if (PagePrivate(page)) {
8889                 ClearPagePrivate(page);
8890                 set_page_private(page, 0);
8891                 put_page(page);
8892         }
8893 }
8894
8895 /*
8896  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8897  * called from a page fault handler when a page is first dirtied. Hence we must
8898  * be careful to check for EOF conditions here. We set the page up correctly
8899  * for a written page which means we get ENOSPC checking when writing into
8900  * holes and correct delalloc and unwritten extent mapping on filesystems that
8901  * support these features.
8902  *
8903  * We are not allowed to take the i_mutex here so we have to play games to
8904  * protect against truncate races as the page could now be beyond EOF.  Because
8905  * vmtruncate() writes the inode size before removing pages, once we have the
8906  * page lock we can determine safely if the page is beyond EOF. If it is not
8907  * beyond EOF, then the page is guaranteed safe against truncation until we
8908  * unlock the page.
8909  */
8910 int btrfs_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
8911 {
8912         struct page *page = vmf->page;
8913         struct inode *inode = file_inode(vma->vm_file);
8914         struct btrfs_root *root = BTRFS_I(inode)->root;
8915         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8916         struct btrfs_ordered_extent *ordered;
8917         struct extent_state *cached_state = NULL;
8918         char *kaddr;
8919         unsigned long zero_start;
8920         loff_t size;
8921         int ret;
8922         int reserved = 0;
8923         u64 reserved_space;
8924         u64 page_start;
8925         u64 page_end;
8926         u64 end;
8927
8928         reserved_space = PAGE_SIZE;
8929
8930         sb_start_pagefault(inode->i_sb);
8931         page_start = page_offset(page);
8932         page_end = page_start + PAGE_SIZE - 1;
8933         end = page_end;
8934
8935         /*
8936          * Reserving delalloc space after obtaining the page lock can lead to
8937          * deadlock. For example, if a dirty page is locked by this function
8938          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8939          * dirty page write out, then the btrfs_writepage() function could
8940          * end up waiting indefinitely to get a lock on the page currently
8941          * being processed by btrfs_page_mkwrite() function.
8942          */
8943         ret = btrfs_delalloc_reserve_space(inode, page_start,
8944                                            reserved_space);
8945         if (!ret) {
8946                 ret = file_update_time(vma->vm_file);
8947                 reserved = 1;
8948         }
8949         if (ret) {
8950                 if (ret == -ENOMEM)
8951                         ret = VM_FAULT_OOM;
8952                 else /* -ENOSPC, -EIO, etc */
8953                         ret = VM_FAULT_SIGBUS;
8954                 if (reserved)
8955                         goto out;
8956                 goto out_noreserve;
8957         }
8958
8959         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8960 again:
8961         lock_page(page);
8962         size = i_size_read(inode);
8963
8964         if ((page->mapping != inode->i_mapping) ||
8965             (page_start >= size)) {
8966                 /* page got truncated out from underneath us */
8967                 goto out_unlock;
8968         }
8969         wait_on_page_writeback(page);
8970
8971         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8972         set_page_extent_mapped(page);
8973
8974         /*
8975          * we can't set the delalloc bits if there are pending ordered
8976          * extents.  Drop our locks and wait for them to finish
8977          */
8978         ordered = btrfs_lookup_ordered_range(inode, page_start, page_end);
8979         if (ordered) {
8980                 unlock_extent_cached(io_tree, page_start, page_end,
8981                                      &cached_state, GFP_NOFS);
8982                 unlock_page(page);
8983                 btrfs_start_ordered_extent(inode, ordered, 1);
8984                 btrfs_put_ordered_extent(ordered);
8985                 goto again;
8986         }
8987
8988         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
8989                 reserved_space = round_up(size - page_start, root->sectorsize);
8990                 if (reserved_space < PAGE_SIZE) {
8991                         end = page_start + reserved_space - 1;
8992                         spin_lock(&BTRFS_I(inode)->lock);
8993                         BTRFS_I(inode)->outstanding_extents++;
8994                         spin_unlock(&BTRFS_I(inode)->lock);
8995                         btrfs_delalloc_release_space(inode, page_start,
8996                                                 PAGE_SIZE - reserved_space);
8997                 }
8998         }
8999
9000         /*
9001          * XXX - page_mkwrite gets called every time the page is dirtied, even
9002          * if it was already dirty, so for space accounting reasons we need to
9003          * clear any delalloc bits for the range we are fixing to save.  There
9004          * is probably a better way to do this, but for now keep consistent with
9005          * prepare_pages in the normal write path.
9006          */
9007         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9008                           EXTENT_DIRTY | EXTENT_DELALLOC |
9009                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9010                           0, 0, &cached_state, GFP_NOFS);
9011
9012         ret = btrfs_set_extent_delalloc(inode, page_start, end,
9013                                         &cached_state);
9014         if (ret) {
9015                 unlock_extent_cached(io_tree, page_start, page_end,
9016                                      &cached_state, GFP_NOFS);
9017                 ret = VM_FAULT_SIGBUS;
9018                 goto out_unlock;
9019         }
9020         ret = 0;
9021
9022         /* page is wholly or partially inside EOF */
9023         if (page_start + PAGE_SIZE > size)
9024                 zero_start = size & ~PAGE_MASK;
9025         else
9026                 zero_start = PAGE_SIZE;
9027
9028         if (zero_start != PAGE_SIZE) {
9029                 kaddr = kmap(page);
9030                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9031                 flush_dcache_page(page);
9032                 kunmap(page);
9033         }
9034         ClearPageChecked(page);
9035         set_page_dirty(page);
9036         SetPageUptodate(page);
9037
9038         BTRFS_I(inode)->last_trans = root->fs_info->generation;
9039         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9040         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9041
9042         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
9043
9044 out_unlock:
9045         if (!ret) {
9046                 sb_end_pagefault(inode->i_sb);
9047                 return VM_FAULT_LOCKED;
9048         }
9049         unlock_page(page);
9050 out:
9051         btrfs_delalloc_release_space(inode, page_start, reserved_space);
9052 out_noreserve:
9053         sb_end_pagefault(inode->i_sb);
9054         return ret;
9055 }
9056
9057 static int btrfs_truncate(struct inode *inode)
9058 {
9059         struct btrfs_root *root = BTRFS_I(inode)->root;
9060         struct btrfs_block_rsv *rsv;
9061         int ret = 0;
9062         int err = 0;
9063         struct btrfs_trans_handle *trans;
9064         u64 mask = root->sectorsize - 1;
9065         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
9066
9067         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9068                                        (u64)-1);
9069         if (ret)
9070                 return ret;
9071
9072         /*
9073          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9074          * 3 things going on here
9075          *
9076          * 1) We need to reserve space for our orphan item and the space to
9077          * delete our orphan item.  Lord knows we don't want to have a dangling
9078          * orphan item because we didn't reserve space to remove it.
9079          *
9080          * 2) We need to reserve space to update our inode.
9081          *
9082          * 3) We need to have something to cache all the space that is going to
9083          * be free'd up by the truncate operation, but also have some slack
9084          * space reserved in case it uses space during the truncate (thank you
9085          * very much snapshotting).
9086          *
9087          * And we need these to all be separate.  The fact is we can use a lot of
9088          * space doing the truncate, and we have no earthly idea how much space
9089          * we will use, so we need the truncate reservation to be separate so it
9090          * doesn't end up using space reserved for updating the inode or
9091          * removing the orphan item.  We also need to be able to stop the
9092          * transaction and start a new one, which means we need to be able to
9093          * update the inode several times, and we have no idea of knowing how
9094          * many times that will be, so we can't just reserve 1 item for the
9095          * entirety of the operation, so that has to be done separately as well.
9096          * Then there is the orphan item, which does indeed need to be held on
9097          * to for the whole operation, and we need nobody to touch this reserved
9098          * space except the orphan code.
9099          *
9100          * So that leaves us with
9101          *
9102          * 1) root->orphan_block_rsv - for the orphan deletion.
9103          * 2) rsv - for the truncate reservation, which we will steal from the
9104          * transaction reservation.
9105          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9106          * updating the inode.
9107          */
9108         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
9109         if (!rsv)
9110                 return -ENOMEM;
9111         rsv->size = min_size;
9112         rsv->failfast = 1;
9113
9114         /*
9115          * 1 for the truncate slack space
9116          * 1 for updating the inode.
9117          */
9118         trans = btrfs_start_transaction(root, 2);
9119         if (IS_ERR(trans)) {
9120                 err = PTR_ERR(trans);
9121                 goto out;
9122         }
9123
9124         /* Migrate the slack space for the truncate to our reserve */
9125         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
9126                                       min_size, 0);
9127         BUG_ON(ret);
9128
9129         /*
9130          * So if we truncate and then write and fsync we normally would just
9131          * write the extents that changed, which is a problem if we need to
9132          * first truncate that entire inode.  So set this flag so we write out
9133          * all of the extents in the inode to the sync log so we're completely
9134          * safe.
9135          */
9136         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9137         trans->block_rsv = rsv;
9138
9139         while (1) {
9140                 ret = btrfs_truncate_inode_items(trans, root, inode,
9141                                                  inode->i_size,
9142                                                  BTRFS_EXTENT_DATA_KEY);
9143                 if (ret != -ENOSPC && ret != -EAGAIN) {
9144                         err = ret;
9145                         break;
9146                 }
9147
9148                 trans->block_rsv = &root->fs_info->trans_block_rsv;
9149                 ret = btrfs_update_inode(trans, root, inode);
9150                 if (ret) {
9151                         err = ret;
9152                         break;
9153                 }
9154
9155                 btrfs_end_transaction(trans, root);
9156                 btrfs_btree_balance_dirty(root);
9157
9158                 trans = btrfs_start_transaction(root, 2);
9159                 if (IS_ERR(trans)) {
9160                         ret = err = PTR_ERR(trans);
9161                         trans = NULL;
9162                         break;
9163                 }
9164
9165                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
9166                                               rsv, min_size, 0);
9167                 BUG_ON(ret);    /* shouldn't happen */
9168                 trans->block_rsv = rsv;
9169         }
9170
9171         if (ret == 0 && inode->i_nlink > 0) {
9172                 trans->block_rsv = root->orphan_block_rsv;
9173                 ret = btrfs_orphan_del(trans, inode);
9174                 if (ret)
9175                         err = ret;
9176         }
9177
9178         if (trans) {
9179                 trans->block_rsv = &root->fs_info->trans_block_rsv;
9180                 ret = btrfs_update_inode(trans, root, inode);
9181                 if (ret && !err)
9182                         err = ret;
9183
9184                 ret = btrfs_end_transaction(trans, root);
9185                 btrfs_btree_balance_dirty(root);
9186         }
9187 out:
9188         btrfs_free_block_rsv(root, rsv);
9189
9190         if (ret && !err)
9191                 err = ret;
9192
9193         return err;
9194 }
9195
9196 /*
9197  * create a new subvolume directory/inode (helper for the ioctl).
9198  */
9199 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9200                              struct btrfs_root *new_root,
9201                              struct btrfs_root *parent_root,
9202                              u64 new_dirid)
9203 {
9204         struct inode *inode;
9205         int err;
9206         u64 index = 0;
9207
9208         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9209                                 new_dirid, new_dirid,
9210                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9211                                 &index);
9212         if (IS_ERR(inode))
9213                 return PTR_ERR(inode);
9214         inode->i_op = &btrfs_dir_inode_operations;
9215         inode->i_fop = &btrfs_dir_file_operations;
9216
9217         set_nlink(inode, 1);
9218         btrfs_i_size_write(inode, 0);
9219         unlock_new_inode(inode);
9220
9221         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9222         if (err)
9223                 btrfs_err(new_root->fs_info,
9224                           "error inheriting subvolume %llu properties: %d",
9225                           new_root->root_key.objectid, err);
9226
9227         err = btrfs_update_inode(trans, new_root, inode);
9228
9229         iput(inode);
9230         return err;
9231 }
9232
9233 struct inode *btrfs_alloc_inode(struct super_block *sb)
9234 {
9235         struct btrfs_inode *ei;
9236         struct inode *inode;
9237
9238         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9239         if (!ei)
9240                 return NULL;
9241
9242         ei->root = NULL;
9243         ei->generation = 0;
9244         ei->last_trans = 0;
9245         ei->last_sub_trans = 0;
9246         ei->logged_trans = 0;
9247         ei->delalloc_bytes = 0;
9248         ei->defrag_bytes = 0;
9249         ei->disk_i_size = 0;
9250         ei->flags = 0;
9251         ei->csum_bytes = 0;
9252         ei->index_cnt = (u64)-1;
9253         ei->dir_index = 0;
9254         ei->last_unlink_trans = 0;
9255         ei->last_log_commit = 0;
9256         ei->delayed_iput_count = 0;
9257
9258         spin_lock_init(&ei->lock);
9259         ei->outstanding_extents = 0;
9260         ei->reserved_extents = 0;
9261
9262         ei->runtime_flags = 0;
9263         ei->force_compress = BTRFS_COMPRESS_NONE;
9264
9265         ei->delayed_node = NULL;
9266
9267         ei->i_otime.tv_sec = 0;
9268         ei->i_otime.tv_nsec = 0;
9269
9270         inode = &ei->vfs_inode;
9271         extent_map_tree_init(&ei->extent_tree);
9272         extent_io_tree_init(&ei->io_tree, &inode->i_data);
9273         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
9274         ei->io_tree.track_uptodate = 1;
9275         ei->io_failure_tree.track_uptodate = 1;
9276         atomic_set(&ei->sync_writers, 0);
9277         mutex_init(&ei->log_mutex);
9278         mutex_init(&ei->delalloc_mutex);
9279         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9280         INIT_LIST_HEAD(&ei->delalloc_inodes);
9281         INIT_LIST_HEAD(&ei->delayed_iput);
9282         RB_CLEAR_NODE(&ei->rb_node);
9283         init_rwsem(&ei->dio_sem);
9284
9285         return inode;
9286 }
9287
9288 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9289 void btrfs_test_destroy_inode(struct inode *inode)
9290 {
9291         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9292         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9293 }
9294 #endif
9295
9296 static void btrfs_i_callback(struct rcu_head *head)
9297 {
9298         struct inode *inode = container_of(head, struct inode, i_rcu);
9299         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9300 }
9301
9302 void btrfs_destroy_inode(struct inode *inode)
9303 {
9304         struct btrfs_ordered_extent *ordered;
9305         struct btrfs_root *root = BTRFS_I(inode)->root;
9306
9307         WARN_ON(!hlist_empty(&inode->i_dentry));
9308         WARN_ON(inode->i_data.nrpages);
9309         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9310         WARN_ON(BTRFS_I(inode)->reserved_extents);
9311         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9312         WARN_ON(BTRFS_I(inode)->csum_bytes);
9313         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9314
9315         /*
9316          * This can happen where we create an inode, but somebody else also
9317          * created the same inode and we need to destroy the one we already
9318          * created.
9319          */
9320         if (!root)
9321                 goto free;
9322
9323         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9324                      &BTRFS_I(inode)->runtime_flags)) {
9325                 btrfs_info(root->fs_info, "inode %llu still on the orphan list",
9326                         btrfs_ino(inode));
9327                 atomic_dec(&root->orphan_inodes);
9328         }
9329
9330         while (1) {
9331                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9332                 if (!ordered)
9333                         break;
9334                 else {
9335                         btrfs_err(root->fs_info, "found ordered extent %llu %llu on inode cleanup",
9336                                 ordered->file_offset, ordered->len);
9337                         btrfs_remove_ordered_extent(inode, ordered);
9338                         btrfs_put_ordered_extent(ordered);
9339                         btrfs_put_ordered_extent(ordered);
9340                 }
9341         }
9342         btrfs_qgroup_check_reserved_leak(inode);
9343         inode_tree_del(inode);
9344         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9345 free:
9346         call_rcu(&inode->i_rcu, btrfs_i_callback);
9347 }
9348
9349 int btrfs_drop_inode(struct inode *inode)
9350 {
9351         struct btrfs_root *root = BTRFS_I(inode)->root;
9352
9353         if (root == NULL)
9354                 return 1;
9355
9356         /* the snap/subvol tree is on deleting */
9357         if (btrfs_root_refs(&root->root_item) == 0)
9358                 return 1;
9359         else
9360                 return generic_drop_inode(inode);
9361 }
9362
9363 static void init_once(void *foo)
9364 {
9365         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9366
9367         inode_init_once(&ei->vfs_inode);
9368 }
9369
9370 void btrfs_destroy_cachep(void)
9371 {
9372         /*
9373          * Make sure all delayed rcu free inodes are flushed before we
9374          * destroy cache.
9375          */
9376         rcu_barrier();
9377         kmem_cache_destroy(btrfs_inode_cachep);
9378         kmem_cache_destroy(btrfs_trans_handle_cachep);
9379         kmem_cache_destroy(btrfs_transaction_cachep);
9380         kmem_cache_destroy(btrfs_path_cachep);
9381         kmem_cache_destroy(btrfs_free_space_cachep);
9382 }
9383
9384 int btrfs_init_cachep(void)
9385 {
9386         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9387                         sizeof(struct btrfs_inode), 0,
9388                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9389                         init_once);
9390         if (!btrfs_inode_cachep)
9391                 goto fail;
9392
9393         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9394                         sizeof(struct btrfs_trans_handle), 0,
9395                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9396         if (!btrfs_trans_handle_cachep)
9397                 goto fail;
9398
9399         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
9400                         sizeof(struct btrfs_transaction), 0,
9401                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9402         if (!btrfs_transaction_cachep)
9403                 goto fail;
9404
9405         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9406                         sizeof(struct btrfs_path), 0,
9407                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9408         if (!btrfs_path_cachep)
9409                 goto fail;
9410
9411         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9412                         sizeof(struct btrfs_free_space), 0,
9413                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9414         if (!btrfs_free_space_cachep)
9415                 goto fail;
9416
9417         return 0;
9418 fail:
9419         btrfs_destroy_cachep();
9420         return -ENOMEM;
9421 }
9422
9423 static int btrfs_getattr(struct vfsmount *mnt,
9424                          struct dentry *dentry, struct kstat *stat)
9425 {
9426         u64 delalloc_bytes;
9427         struct inode *inode = d_inode(dentry);
9428         u32 blocksize = inode->i_sb->s_blocksize;
9429
9430         generic_fillattr(inode, stat);
9431         stat->dev = BTRFS_I(inode)->root->anon_dev;
9432
9433         spin_lock(&BTRFS_I(inode)->lock);
9434         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
9435         spin_unlock(&BTRFS_I(inode)->lock);
9436         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9437                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9438         return 0;
9439 }
9440
9441 static int btrfs_rename_exchange(struct inode *old_dir,
9442                               struct dentry *old_dentry,
9443                               struct inode *new_dir,
9444                               struct dentry *new_dentry)
9445 {
9446         struct btrfs_trans_handle *trans;
9447         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9448         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9449         struct inode *new_inode = new_dentry->d_inode;
9450         struct inode *old_inode = old_dentry->d_inode;
9451         struct timespec ctime = CURRENT_TIME;
9452         struct dentry *parent;
9453         u64 old_ino = btrfs_ino(old_inode);
9454         u64 new_ino = btrfs_ino(new_inode);
9455         u64 old_idx = 0;
9456         u64 new_idx = 0;
9457         u64 root_objectid;
9458         int ret;
9459         bool root_log_pinned = false;
9460         bool dest_log_pinned = false;
9461
9462         /* we only allow rename subvolume link between subvolumes */
9463         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9464                 return -EXDEV;
9465
9466         /* close the race window with snapshot create/destroy ioctl */
9467         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9468                 down_read(&root->fs_info->subvol_sem);
9469         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9470                 down_read(&dest->fs_info->subvol_sem);
9471
9472         /*
9473          * We want to reserve the absolute worst case amount of items.  So if
9474          * both inodes are subvols and we need to unlink them then that would
9475          * require 4 item modifications, but if they are both normal inodes it
9476          * would require 5 item modifications, so we'll assume their normal
9477          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9478          * should cover the worst case number of items we'll modify.
9479          */
9480         trans = btrfs_start_transaction(root, 12);
9481         if (IS_ERR(trans)) {
9482                 ret = PTR_ERR(trans);
9483                 goto out_notrans;
9484         }
9485
9486         /*
9487          * We need to find a free sequence number both in the source and
9488          * in the destination directory for the exchange.
9489          */
9490         ret = btrfs_set_inode_index(new_dir, &old_idx);
9491         if (ret)
9492                 goto out_fail;
9493         ret = btrfs_set_inode_index(old_dir, &new_idx);
9494         if (ret)
9495                 goto out_fail;
9496
9497         BTRFS_I(old_inode)->dir_index = 0ULL;
9498         BTRFS_I(new_inode)->dir_index = 0ULL;
9499
9500         /* Reference for the source. */
9501         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9502                 /* force full log commit if subvolume involved. */
9503                 btrfs_set_log_full_commit(root->fs_info, trans);
9504         } else {
9505                 btrfs_pin_log_trans(root);
9506                 root_log_pinned = true;
9507                 ret = btrfs_insert_inode_ref(trans, dest,
9508                                              new_dentry->d_name.name,
9509                                              new_dentry->d_name.len,
9510                                              old_ino,
9511                                              btrfs_ino(new_dir), old_idx);
9512                 if (ret)
9513                         goto out_fail;
9514         }
9515
9516         /* And now for the dest. */
9517         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9518                 /* force full log commit if subvolume involved. */
9519                 btrfs_set_log_full_commit(dest->fs_info, trans);
9520         } else {
9521                 btrfs_pin_log_trans(dest);
9522                 dest_log_pinned = true;
9523                 ret = btrfs_insert_inode_ref(trans, root,
9524                                              old_dentry->d_name.name,
9525                                              old_dentry->d_name.len,
9526                                              new_ino,
9527                                              btrfs_ino(old_dir), new_idx);
9528                 if (ret)
9529                         goto out_fail;
9530         }
9531
9532         /* Update inode version and ctime/mtime. */
9533         inode_inc_iversion(old_dir);
9534         inode_inc_iversion(new_dir);
9535         inode_inc_iversion(old_inode);
9536         inode_inc_iversion(new_inode);
9537         old_dir->i_ctime = old_dir->i_mtime = ctime;
9538         new_dir->i_ctime = new_dir->i_mtime = ctime;
9539         old_inode->i_ctime = ctime;
9540         new_inode->i_ctime = ctime;
9541
9542         if (old_dentry->d_parent != new_dentry->d_parent) {
9543                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9544                 btrfs_record_unlink_dir(trans, new_dir, new_inode, 1);
9545         }
9546
9547         /* src is a subvolume */
9548         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9549                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9550                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9551                                           root_objectid,
9552                                           old_dentry->d_name.name,
9553                                           old_dentry->d_name.len);
9554         } else { /* src is an inode */
9555                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9556                                            old_dentry->d_inode,
9557                                            old_dentry->d_name.name,
9558                                            old_dentry->d_name.len);
9559                 if (!ret)
9560                         ret = btrfs_update_inode(trans, root, old_inode);
9561         }
9562         if (ret) {
9563                 btrfs_abort_transaction(trans, root, ret);
9564                 goto out_fail;
9565         }
9566
9567         /* dest is a subvolume */
9568         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9569                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9570                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9571                                           root_objectid,
9572                                           new_dentry->d_name.name,
9573                                           new_dentry->d_name.len);
9574         } else { /* dest is an inode */
9575                 ret = __btrfs_unlink_inode(trans, dest, new_dir,
9576                                            new_dentry->d_inode,
9577                                            new_dentry->d_name.name,
9578                                            new_dentry->d_name.len);
9579                 if (!ret)
9580                         ret = btrfs_update_inode(trans, dest, new_inode);
9581         }
9582         if (ret) {
9583                 btrfs_abort_transaction(trans, root, ret);
9584                 goto out_fail;
9585         }
9586
9587         ret = btrfs_add_link(trans, new_dir, old_inode,
9588                              new_dentry->d_name.name,
9589                              new_dentry->d_name.len, 0, old_idx);
9590         if (ret) {
9591                 btrfs_abort_transaction(trans, root, ret);
9592                 goto out_fail;
9593         }
9594
9595         ret = btrfs_add_link(trans, old_dir, new_inode,
9596                              old_dentry->d_name.name,
9597                              old_dentry->d_name.len, 0, new_idx);
9598         if (ret) {
9599                 btrfs_abort_transaction(trans, root, ret);
9600                 goto out_fail;
9601         }
9602
9603         if (old_inode->i_nlink == 1)
9604                 BTRFS_I(old_inode)->dir_index = old_idx;
9605         if (new_inode->i_nlink == 1)
9606                 BTRFS_I(new_inode)->dir_index = new_idx;
9607
9608         if (root_log_pinned) {
9609                 parent = new_dentry->d_parent;
9610                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9611                 btrfs_end_log_trans(root);
9612                 root_log_pinned = false;
9613         }
9614         if (dest_log_pinned) {
9615                 parent = old_dentry->d_parent;
9616                 btrfs_log_new_name(trans, new_inode, new_dir, parent);
9617                 btrfs_end_log_trans(dest);
9618                 dest_log_pinned = false;
9619         }
9620 out_fail:
9621         /*
9622          * If we have pinned a log and an error happened, we unpin tasks
9623          * trying to sync the log and force them to fallback to a transaction
9624          * commit if the log currently contains any of the inodes involved in
9625          * this rename operation (to ensure we do not persist a log with an
9626          * inconsistent state for any of these inodes or leading to any
9627          * inconsistencies when replayed). If the transaction was aborted, the
9628          * abortion reason is propagated to userspace when attempting to commit
9629          * the transaction. If the log does not contain any of these inodes, we
9630          * allow the tasks to sync it.
9631          */
9632         if (ret && (root_log_pinned || dest_log_pinned)) {
9633                 if (btrfs_inode_in_log(old_dir, root->fs_info->generation) ||
9634                     btrfs_inode_in_log(new_dir, root->fs_info->generation) ||
9635                     btrfs_inode_in_log(old_inode, root->fs_info->generation) ||
9636                     (new_inode &&
9637                      btrfs_inode_in_log(new_inode, root->fs_info->generation)))
9638                     btrfs_set_log_full_commit(root->fs_info, trans);
9639
9640                 if (root_log_pinned) {
9641                         btrfs_end_log_trans(root);
9642                         root_log_pinned = false;
9643                 }
9644                 if (dest_log_pinned) {
9645                         btrfs_end_log_trans(dest);
9646                         dest_log_pinned = false;
9647                 }
9648         }
9649         ret = btrfs_end_transaction(trans, root);
9650 out_notrans:
9651         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9652                 up_read(&dest->fs_info->subvol_sem);
9653         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9654                 up_read(&root->fs_info->subvol_sem);
9655
9656         return ret;
9657 }
9658
9659 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9660                                      struct btrfs_root *root,
9661                                      struct inode *dir,
9662                                      struct dentry *dentry)
9663 {
9664         int ret;
9665         struct inode *inode;
9666         u64 objectid;
9667         u64 index;
9668
9669         ret = btrfs_find_free_ino(root, &objectid);
9670         if (ret)
9671                 return ret;
9672
9673         inode = btrfs_new_inode(trans, root, dir,
9674                                 dentry->d_name.name,
9675                                 dentry->d_name.len,
9676                                 btrfs_ino(dir),
9677                                 objectid,
9678                                 S_IFCHR | WHITEOUT_MODE,
9679                                 &index);
9680
9681         if (IS_ERR(inode)) {
9682                 ret = PTR_ERR(inode);
9683                 return ret;
9684         }
9685
9686         inode->i_op = &btrfs_special_inode_operations;
9687         init_special_inode(inode, inode->i_mode,
9688                 WHITEOUT_DEV);
9689
9690         ret = btrfs_init_inode_security(trans, inode, dir,
9691                                 &dentry->d_name);
9692         if (ret)
9693                 goto out;
9694
9695         ret = btrfs_add_nondir(trans, dir, dentry,
9696                                 inode, 0, index);
9697         if (ret)
9698                 goto out;
9699
9700         ret = btrfs_update_inode(trans, root, inode);
9701 out:
9702         unlock_new_inode(inode);
9703         if (ret)
9704                 inode_dec_link_count(inode);
9705         iput(inode);
9706
9707         return ret;
9708 }
9709
9710 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9711                            struct inode *new_dir, struct dentry *new_dentry,
9712                            unsigned int flags)
9713 {
9714         struct btrfs_trans_handle *trans;
9715         unsigned int trans_num_items;
9716         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9717         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9718         struct inode *new_inode = d_inode(new_dentry);
9719         struct inode *old_inode = d_inode(old_dentry);
9720         u64 index = 0;
9721         u64 root_objectid;
9722         int ret;
9723         u64 old_ino = btrfs_ino(old_inode);
9724         bool log_pinned = false;
9725
9726         if (btrfs_ino(new_dir) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9727                 return -EPERM;
9728
9729         /* we only allow rename subvolume link between subvolumes */
9730         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9731                 return -EXDEV;
9732
9733         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9734             (new_inode && btrfs_ino(new_inode) == BTRFS_FIRST_FREE_OBJECTID))
9735                 return -ENOTEMPTY;
9736
9737         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9738             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9739                 return -ENOTEMPTY;
9740
9741
9742         /* check for collisions, even if the  name isn't there */
9743         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9744                              new_dentry->d_name.name,
9745                              new_dentry->d_name.len);
9746
9747         if (ret) {
9748                 if (ret == -EEXIST) {
9749                         /* we shouldn't get
9750                          * eexist without a new_inode */
9751                         if (WARN_ON(!new_inode)) {
9752                                 return ret;
9753                         }
9754                 } else {
9755                         /* maybe -EOVERFLOW */
9756                         return ret;
9757                 }
9758         }
9759         ret = 0;
9760
9761         /*
9762          * we're using rename to replace one file with another.  Start IO on it
9763          * now so  we don't add too much work to the end of the transaction
9764          */
9765         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9766                 filemap_flush(old_inode->i_mapping);
9767
9768         /* close the racy window with snapshot create/destroy ioctl */
9769         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9770                 down_read(&root->fs_info->subvol_sem);
9771         /*
9772          * We want to reserve the absolute worst case amount of items.  So if
9773          * both inodes are subvols and we need to unlink them then that would
9774          * require 4 item modifications, but if they are both normal inodes it
9775          * would require 5 item modifications, so we'll assume they are normal
9776          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9777          * should cover the worst case number of items we'll modify.
9778          * If our rename has the whiteout flag, we need more 5 units for the
9779          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9780          * when selinux is enabled).
9781          */
9782         trans_num_items = 11;
9783         if (flags & RENAME_WHITEOUT)
9784                 trans_num_items += 5;
9785         trans = btrfs_start_transaction(root, trans_num_items);
9786         if (IS_ERR(trans)) {
9787                 ret = PTR_ERR(trans);
9788                 goto out_notrans;
9789         }
9790
9791         if (dest != root)
9792                 btrfs_record_root_in_trans(trans, dest);
9793
9794         ret = btrfs_set_inode_index(new_dir, &index);
9795         if (ret)
9796                 goto out_fail;
9797
9798         BTRFS_I(old_inode)->dir_index = 0ULL;
9799         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9800                 /* force full log commit if subvolume involved. */
9801                 btrfs_set_log_full_commit(root->fs_info, trans);
9802         } else {
9803                 btrfs_pin_log_trans(root);
9804                 log_pinned = true;
9805                 ret = btrfs_insert_inode_ref(trans, dest,
9806                                              new_dentry->d_name.name,
9807                                              new_dentry->d_name.len,
9808                                              old_ino,
9809                                              btrfs_ino(new_dir), index);
9810                 if (ret)
9811                         goto out_fail;
9812         }
9813
9814         inode_inc_iversion(old_dir);
9815         inode_inc_iversion(new_dir);
9816         inode_inc_iversion(old_inode);
9817         old_dir->i_ctime = old_dir->i_mtime =
9818         new_dir->i_ctime = new_dir->i_mtime =
9819         old_inode->i_ctime = current_fs_time(old_dir->i_sb);
9820
9821         if (old_dentry->d_parent != new_dentry->d_parent)
9822                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9823
9824         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9825                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9826                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9827                                         old_dentry->d_name.name,
9828                                         old_dentry->d_name.len);
9829         } else {
9830                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9831                                         d_inode(old_dentry),
9832                                         old_dentry->d_name.name,
9833                                         old_dentry->d_name.len);
9834                 if (!ret)
9835                         ret = btrfs_update_inode(trans, root, old_inode);
9836         }
9837         if (ret) {
9838                 btrfs_abort_transaction(trans, root, ret);
9839                 goto out_fail;
9840         }
9841
9842         if (new_inode) {
9843                 inode_inc_iversion(new_inode);
9844                 new_inode->i_ctime = current_fs_time(new_inode->i_sb);
9845                 if (unlikely(btrfs_ino(new_inode) ==
9846                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9847                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9848                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9849                                                 root_objectid,
9850                                                 new_dentry->d_name.name,
9851                                                 new_dentry->d_name.len);
9852                         BUG_ON(new_inode->i_nlink == 0);
9853                 } else {
9854                         ret = btrfs_unlink_inode(trans, dest, new_dir,
9855                                                  d_inode(new_dentry),
9856                                                  new_dentry->d_name.name,
9857                                                  new_dentry->d_name.len);
9858                 }
9859                 if (!ret && new_inode->i_nlink == 0)
9860                         ret = btrfs_orphan_add(trans, d_inode(new_dentry));
9861                 if (ret) {
9862                         btrfs_abort_transaction(trans, root, ret);
9863                         goto out_fail;
9864                 }
9865         }
9866
9867         ret = btrfs_add_link(trans, new_dir, old_inode,
9868                              new_dentry->d_name.name,
9869                              new_dentry->d_name.len, 0, index);
9870         if (ret) {
9871                 btrfs_abort_transaction(trans, root, ret);
9872                 goto out_fail;
9873         }
9874
9875         if (old_inode->i_nlink == 1)
9876                 BTRFS_I(old_inode)->dir_index = index;
9877
9878         if (log_pinned) {
9879                 struct dentry *parent = new_dentry->d_parent;
9880
9881                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9882                 btrfs_end_log_trans(root);
9883                 log_pinned = false;
9884         }
9885
9886         if (flags & RENAME_WHITEOUT) {
9887                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9888                                                 old_dentry);
9889
9890                 if (ret) {
9891                         btrfs_abort_transaction(trans, root, ret);
9892                         goto out_fail;
9893                 }
9894         }
9895 out_fail:
9896         /*
9897          * If we have pinned the log and an error happened, we unpin tasks
9898          * trying to sync the log and force them to fallback to a transaction
9899          * commit if the log currently contains any of the inodes involved in
9900          * this rename operation (to ensure we do not persist a log with an
9901          * inconsistent state for any of these inodes or leading to any
9902          * inconsistencies when replayed). If the transaction was aborted, the
9903          * abortion reason is propagated to userspace when attempting to commit
9904          * the transaction. If the log does not contain any of these inodes, we
9905          * allow the tasks to sync it.
9906          */
9907         if (ret && log_pinned) {
9908                 if (btrfs_inode_in_log(old_dir, root->fs_info->generation) ||
9909                     btrfs_inode_in_log(new_dir, root->fs_info->generation) ||
9910                     btrfs_inode_in_log(old_inode, root->fs_info->generation) ||
9911                     (new_inode &&
9912                      btrfs_inode_in_log(new_inode, root->fs_info->generation)))
9913                     btrfs_set_log_full_commit(root->fs_info, trans);
9914
9915                 btrfs_end_log_trans(root);
9916                 log_pinned = false;
9917         }
9918         btrfs_end_transaction(trans, root);
9919 out_notrans:
9920         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9921                 up_read(&root->fs_info->subvol_sem);
9922
9923         return ret;
9924 }
9925
9926 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9927                          struct inode *new_dir, struct dentry *new_dentry,
9928                          unsigned int flags)
9929 {
9930         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
9931                 return -EINVAL;
9932
9933         if (flags & RENAME_EXCHANGE)
9934                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
9935                                           new_dentry);
9936
9937         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
9938 }
9939
9940 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9941 {
9942         struct btrfs_delalloc_work *delalloc_work;
9943         struct inode *inode;
9944
9945         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9946                                      work);
9947         inode = delalloc_work->inode;
9948         filemap_flush(inode->i_mapping);
9949         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9950                                 &BTRFS_I(inode)->runtime_flags))
9951                 filemap_flush(inode->i_mapping);
9952
9953         if (delalloc_work->delay_iput)
9954                 btrfs_add_delayed_iput(inode);
9955         else
9956                 iput(inode);
9957         complete(&delalloc_work->completion);
9958 }
9959
9960 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
9961                                                     int delay_iput)
9962 {
9963         struct btrfs_delalloc_work *work;
9964
9965         work = kmalloc(sizeof(*work), GFP_NOFS);
9966         if (!work)
9967                 return NULL;
9968
9969         init_completion(&work->completion);
9970         INIT_LIST_HEAD(&work->list);
9971         work->inode = inode;
9972         work->delay_iput = delay_iput;
9973         WARN_ON_ONCE(!inode);
9974         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9975                         btrfs_run_delalloc_work, NULL, NULL);
9976
9977         return work;
9978 }
9979
9980 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
9981 {
9982         wait_for_completion(&work->completion);
9983         kfree(work);
9984 }
9985
9986 /*
9987  * some fairly slow code that needs optimization. This walks the list
9988  * of all the inodes with pending delalloc and forces them to disk.
9989  */
9990 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
9991                                    int nr)
9992 {
9993         struct btrfs_inode *binode;
9994         struct inode *inode;
9995         struct btrfs_delalloc_work *work, *next;
9996         struct list_head works;
9997         struct list_head splice;
9998         int ret = 0;
9999
10000         INIT_LIST_HEAD(&works);
10001         INIT_LIST_HEAD(&splice);
10002
10003         mutex_lock(&root->delalloc_mutex);
10004         spin_lock(&root->delalloc_lock);
10005         list_splice_init(&root->delalloc_inodes, &splice);
10006         while (!list_empty(&splice)) {
10007                 binode = list_entry(splice.next, struct btrfs_inode,
10008                                     delalloc_inodes);
10009
10010                 list_move_tail(&binode->delalloc_inodes,
10011                                &root->delalloc_inodes);
10012                 inode = igrab(&binode->vfs_inode);
10013                 if (!inode) {
10014                         cond_resched_lock(&root->delalloc_lock);
10015                         continue;
10016                 }
10017                 spin_unlock(&root->delalloc_lock);
10018
10019                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10020                 if (!work) {
10021                         if (delay_iput)
10022                                 btrfs_add_delayed_iput(inode);
10023                         else
10024                                 iput(inode);
10025                         ret = -ENOMEM;
10026                         goto out;
10027                 }
10028                 list_add_tail(&work->list, &works);
10029                 btrfs_queue_work(root->fs_info->flush_workers,
10030                                  &work->work);
10031                 ret++;
10032                 if (nr != -1 && ret >= nr)
10033                         goto out;
10034                 cond_resched();
10035                 spin_lock(&root->delalloc_lock);
10036         }
10037         spin_unlock(&root->delalloc_lock);
10038
10039 out:
10040         list_for_each_entry_safe(work, next, &works, list) {
10041                 list_del_init(&work->list);
10042                 btrfs_wait_and_free_delalloc_work(work);
10043         }
10044
10045         if (!list_empty_careful(&splice)) {
10046                 spin_lock(&root->delalloc_lock);
10047                 list_splice_tail(&splice, &root->delalloc_inodes);
10048                 spin_unlock(&root->delalloc_lock);
10049         }
10050         mutex_unlock(&root->delalloc_mutex);
10051         return ret;
10052 }
10053
10054 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10055 {
10056         int ret;
10057
10058         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
10059                 return -EROFS;
10060
10061         ret = __start_delalloc_inodes(root, delay_iput, -1);
10062         if (ret > 0)
10063                 ret = 0;
10064         /*
10065          * the filemap_flush will queue IO into the worker threads, but
10066          * we have to make sure the IO is actually started and that
10067          * ordered extents get created before we return
10068          */
10069         atomic_inc(&root->fs_info->async_submit_draining);
10070         while (atomic_read(&root->fs_info->nr_async_submits) ||
10071               atomic_read(&root->fs_info->async_delalloc_pages)) {
10072                 wait_event(root->fs_info->async_submit_wait,
10073                    (atomic_read(&root->fs_info->nr_async_submits) == 0 &&
10074                     atomic_read(&root->fs_info->async_delalloc_pages) == 0));
10075         }
10076         atomic_dec(&root->fs_info->async_submit_draining);
10077         return ret;
10078 }
10079
10080 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10081                                int nr)
10082 {
10083         struct btrfs_root *root;
10084         struct list_head splice;
10085         int ret;
10086
10087         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10088                 return -EROFS;
10089
10090         INIT_LIST_HEAD(&splice);
10091
10092         mutex_lock(&fs_info->delalloc_root_mutex);
10093         spin_lock(&fs_info->delalloc_root_lock);
10094         list_splice_init(&fs_info->delalloc_roots, &splice);
10095         while (!list_empty(&splice) && nr) {
10096                 root = list_first_entry(&splice, struct btrfs_root,
10097                                         delalloc_root);
10098                 root = btrfs_grab_fs_root(root);
10099                 BUG_ON(!root);
10100                 list_move_tail(&root->delalloc_root,
10101                                &fs_info->delalloc_roots);
10102                 spin_unlock(&fs_info->delalloc_root_lock);
10103
10104                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10105                 btrfs_put_fs_root(root);
10106                 if (ret < 0)
10107                         goto out;
10108
10109                 if (nr != -1) {
10110                         nr -= ret;
10111                         WARN_ON(nr < 0);
10112                 }
10113                 spin_lock(&fs_info->delalloc_root_lock);
10114         }
10115         spin_unlock(&fs_info->delalloc_root_lock);
10116
10117         ret = 0;
10118         atomic_inc(&fs_info->async_submit_draining);
10119         while (atomic_read(&fs_info->nr_async_submits) ||
10120               atomic_read(&fs_info->async_delalloc_pages)) {
10121                 wait_event(fs_info->async_submit_wait,
10122                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
10123                     atomic_read(&fs_info->async_delalloc_pages) == 0));
10124         }
10125         atomic_dec(&fs_info->async_submit_draining);
10126 out:
10127         if (!list_empty_careful(&splice)) {
10128                 spin_lock(&fs_info->delalloc_root_lock);
10129                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10130                 spin_unlock(&fs_info->delalloc_root_lock);
10131         }
10132         mutex_unlock(&fs_info->delalloc_root_mutex);
10133         return ret;
10134 }
10135
10136 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10137                          const char *symname)
10138 {
10139         struct btrfs_trans_handle *trans;
10140         struct btrfs_root *root = BTRFS_I(dir)->root;
10141         struct btrfs_path *path;
10142         struct btrfs_key key;
10143         struct inode *inode = NULL;
10144         int err;
10145         int drop_inode = 0;
10146         u64 objectid;
10147         u64 index = 0;
10148         int name_len;
10149         int datasize;
10150         unsigned long ptr;
10151         struct btrfs_file_extent_item *ei;
10152         struct extent_buffer *leaf;
10153
10154         name_len = strlen(symname);
10155         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(root))
10156                 return -ENAMETOOLONG;
10157
10158         /*
10159          * 2 items for inode item and ref
10160          * 2 items for dir items
10161          * 1 item for updating parent inode item
10162          * 1 item for the inline extent item
10163          * 1 item for xattr if selinux is on
10164          */
10165         trans = btrfs_start_transaction(root, 7);
10166         if (IS_ERR(trans))
10167                 return PTR_ERR(trans);
10168
10169         err = btrfs_find_free_ino(root, &objectid);
10170         if (err)
10171                 goto out_unlock;
10172
10173         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10174                                 dentry->d_name.len, btrfs_ino(dir), objectid,
10175                                 S_IFLNK|S_IRWXUGO, &index);
10176         if (IS_ERR(inode)) {
10177                 err = PTR_ERR(inode);
10178                 goto out_unlock;
10179         }
10180
10181         /*
10182         * If the active LSM wants to access the inode during
10183         * d_instantiate it needs these. Smack checks to see
10184         * if the filesystem supports xattrs by looking at the
10185         * ops vector.
10186         */
10187         inode->i_fop = &btrfs_file_operations;
10188         inode->i_op = &btrfs_file_inode_operations;
10189         inode->i_mapping->a_ops = &btrfs_aops;
10190         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10191
10192         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10193         if (err)
10194                 goto out_unlock_inode;
10195
10196         path = btrfs_alloc_path();
10197         if (!path) {
10198                 err = -ENOMEM;
10199                 goto out_unlock_inode;
10200         }
10201         key.objectid = btrfs_ino(inode);
10202         key.offset = 0;
10203         key.type = BTRFS_EXTENT_DATA_KEY;
10204         datasize = btrfs_file_extent_calc_inline_size(name_len);
10205         err = btrfs_insert_empty_item(trans, root, path, &key,
10206                                       datasize);
10207         if (err) {
10208                 btrfs_free_path(path);
10209                 goto out_unlock_inode;
10210         }
10211         leaf = path->nodes[0];
10212         ei = btrfs_item_ptr(leaf, path->slots[0],
10213                             struct btrfs_file_extent_item);
10214         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10215         btrfs_set_file_extent_type(leaf, ei,
10216                                    BTRFS_FILE_EXTENT_INLINE);
10217         btrfs_set_file_extent_encryption(leaf, ei, 0);
10218         btrfs_set_file_extent_compression(leaf, ei, 0);
10219         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10220         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10221
10222         ptr = btrfs_file_extent_inline_start(ei);
10223         write_extent_buffer(leaf, symname, ptr, name_len);
10224         btrfs_mark_buffer_dirty(leaf);
10225         btrfs_free_path(path);
10226
10227         inode->i_op = &btrfs_symlink_inode_operations;
10228         inode_nohighmem(inode);
10229         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10230         inode_set_bytes(inode, name_len);
10231         btrfs_i_size_write(inode, name_len);
10232         err = btrfs_update_inode(trans, root, inode);
10233         /*
10234          * Last step, add directory indexes for our symlink inode. This is the
10235          * last step to avoid extra cleanup of these indexes if an error happens
10236          * elsewhere above.
10237          */
10238         if (!err)
10239                 err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
10240         if (err) {
10241                 drop_inode = 1;
10242                 goto out_unlock_inode;
10243         }
10244
10245         unlock_new_inode(inode);
10246         d_instantiate(dentry, inode);
10247
10248 out_unlock:
10249         btrfs_end_transaction(trans, root);
10250         if (drop_inode) {
10251                 inode_dec_link_count(inode);
10252                 iput(inode);
10253         }
10254         btrfs_btree_balance_dirty(root);
10255         return err;
10256
10257 out_unlock_inode:
10258         drop_inode = 1;
10259         unlock_new_inode(inode);
10260         goto out_unlock;
10261 }
10262
10263 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10264                                        u64 start, u64 num_bytes, u64 min_size,
10265                                        loff_t actual_len, u64 *alloc_hint,
10266                                        struct btrfs_trans_handle *trans)
10267 {
10268         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10269         struct extent_map *em;
10270         struct btrfs_root *root = BTRFS_I(inode)->root;
10271         struct btrfs_key ins;
10272         u64 cur_offset = start;
10273         u64 i_size;
10274         u64 cur_bytes;
10275         u64 last_alloc = (u64)-1;
10276         int ret = 0;
10277         bool own_trans = true;
10278
10279         if (trans)
10280                 own_trans = false;
10281         while (num_bytes > 0) {
10282                 if (own_trans) {
10283                         trans = btrfs_start_transaction(root, 3);
10284                         if (IS_ERR(trans)) {
10285                                 ret = PTR_ERR(trans);
10286                                 break;
10287                         }
10288                 }
10289
10290                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10291                 cur_bytes = max(cur_bytes, min_size);
10292                 /*
10293                  * If we are severely fragmented we could end up with really
10294                  * small allocations, so if the allocator is returning small
10295                  * chunks lets make its job easier by only searching for those
10296                  * sized chunks.
10297                  */
10298                 cur_bytes = min(cur_bytes, last_alloc);
10299                 ret = btrfs_reserve_extent(root, cur_bytes, min_size, 0,
10300                                            *alloc_hint, &ins, 1, 0);
10301                 if (ret) {
10302                         if (own_trans)
10303                                 btrfs_end_transaction(trans, root);
10304                         break;
10305                 }
10306                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
10307
10308                 last_alloc = ins.offset;
10309                 ret = insert_reserved_file_extent(trans, inode,
10310                                                   cur_offset, ins.objectid,
10311                                                   ins.offset, ins.offset,
10312                                                   ins.offset, 0, 0, 0,
10313                                                   BTRFS_FILE_EXTENT_PREALLOC);
10314                 if (ret) {
10315                         btrfs_free_reserved_extent(root, ins.objectid,
10316                                                    ins.offset, 0);
10317                         btrfs_abort_transaction(trans, root, ret);
10318                         if (own_trans)
10319                                 btrfs_end_transaction(trans, root);
10320                         break;
10321                 }
10322
10323                 btrfs_drop_extent_cache(inode, cur_offset,
10324                                         cur_offset + ins.offset -1, 0);
10325
10326                 em = alloc_extent_map();
10327                 if (!em) {
10328                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10329                                 &BTRFS_I(inode)->runtime_flags);
10330                         goto next;
10331                 }
10332
10333                 em->start = cur_offset;
10334                 em->orig_start = cur_offset;
10335                 em->len = ins.offset;
10336                 em->block_start = ins.objectid;
10337                 em->block_len = ins.offset;
10338                 em->orig_block_len = ins.offset;
10339                 em->ram_bytes = ins.offset;
10340                 em->bdev = root->fs_info->fs_devices->latest_bdev;
10341                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10342                 em->generation = trans->transid;
10343
10344                 while (1) {
10345                         write_lock(&em_tree->lock);
10346                         ret = add_extent_mapping(em_tree, em, 1);
10347                         write_unlock(&em_tree->lock);
10348                         if (ret != -EEXIST)
10349                                 break;
10350                         btrfs_drop_extent_cache(inode, cur_offset,
10351                                                 cur_offset + ins.offset - 1,
10352                                                 0);
10353                 }
10354                 free_extent_map(em);
10355 next:
10356                 num_bytes -= ins.offset;
10357                 cur_offset += ins.offset;
10358                 *alloc_hint = ins.objectid + ins.offset;
10359
10360                 inode_inc_iversion(inode);
10361                 inode->i_ctime = current_fs_time(inode->i_sb);
10362                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10363                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10364                     (actual_len > inode->i_size) &&
10365                     (cur_offset > inode->i_size)) {
10366                         if (cur_offset > actual_len)
10367                                 i_size = actual_len;
10368                         else
10369                                 i_size = cur_offset;
10370                         i_size_write(inode, i_size);
10371                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10372                 }
10373
10374                 ret = btrfs_update_inode(trans, root, inode);
10375
10376                 if (ret) {
10377                         btrfs_abort_transaction(trans, root, ret);
10378                         if (own_trans)
10379                                 btrfs_end_transaction(trans, root);
10380                         break;
10381                 }
10382
10383                 if (own_trans)
10384                         btrfs_end_transaction(trans, root);
10385         }
10386         return ret;
10387 }
10388
10389 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10390                               u64 start, u64 num_bytes, u64 min_size,
10391                               loff_t actual_len, u64 *alloc_hint)
10392 {
10393         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10394                                            min_size, actual_len, alloc_hint,
10395                                            NULL);
10396 }
10397
10398 int btrfs_prealloc_file_range_trans(struct inode *inode,
10399                                     struct btrfs_trans_handle *trans, int mode,
10400                                     u64 start, u64 num_bytes, u64 min_size,
10401                                     loff_t actual_len, u64 *alloc_hint)
10402 {
10403         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10404                                            min_size, actual_len, alloc_hint, trans);
10405 }
10406
10407 static int btrfs_set_page_dirty(struct page *page)
10408 {
10409         return __set_page_dirty_nobuffers(page);
10410 }
10411
10412 static int btrfs_permission(struct inode *inode, int mask)
10413 {
10414         struct btrfs_root *root = BTRFS_I(inode)->root;
10415         umode_t mode = inode->i_mode;
10416
10417         if (mask & MAY_WRITE &&
10418             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10419                 if (btrfs_root_readonly(root))
10420                         return -EROFS;
10421                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10422                         return -EACCES;
10423         }
10424         return generic_permission(inode, mask);
10425 }
10426
10427 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10428 {
10429         struct btrfs_trans_handle *trans;
10430         struct btrfs_root *root = BTRFS_I(dir)->root;
10431         struct inode *inode = NULL;
10432         u64 objectid;
10433         u64 index;
10434         int ret = 0;
10435
10436         /*
10437          * 5 units required for adding orphan entry
10438          */
10439         trans = btrfs_start_transaction(root, 5);
10440         if (IS_ERR(trans))
10441                 return PTR_ERR(trans);
10442
10443         ret = btrfs_find_free_ino(root, &objectid);
10444         if (ret)
10445                 goto out;
10446
10447         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10448                                 btrfs_ino(dir), objectid, mode, &index);
10449         if (IS_ERR(inode)) {
10450                 ret = PTR_ERR(inode);
10451                 inode = NULL;
10452                 goto out;
10453         }
10454
10455         inode->i_fop = &btrfs_file_operations;
10456         inode->i_op = &btrfs_file_inode_operations;
10457
10458         inode->i_mapping->a_ops = &btrfs_aops;
10459         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10460
10461         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10462         if (ret)
10463                 goto out_inode;
10464
10465         ret = btrfs_update_inode(trans, root, inode);
10466         if (ret)
10467                 goto out_inode;
10468         ret = btrfs_orphan_add(trans, inode);
10469         if (ret)
10470                 goto out_inode;
10471
10472         /*
10473          * We set number of links to 0 in btrfs_new_inode(), and here we set
10474          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10475          * through:
10476          *
10477          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10478          */
10479         set_nlink(inode, 1);
10480         unlock_new_inode(inode);
10481         d_tmpfile(dentry, inode);
10482         mark_inode_dirty(inode);
10483
10484 out:
10485         btrfs_end_transaction(trans, root);
10486         if (ret)
10487                 iput(inode);
10488         btrfs_balance_delayed_items(root);
10489         btrfs_btree_balance_dirty(root);
10490         return ret;
10491
10492 out_inode:
10493         unlock_new_inode(inode);
10494         goto out;
10495
10496 }
10497
10498 /* Inspired by filemap_check_errors() */
10499 int btrfs_inode_check_errors(struct inode *inode)
10500 {
10501         int ret = 0;
10502
10503         if (test_bit(AS_ENOSPC, &inode->i_mapping->flags) &&
10504             test_and_clear_bit(AS_ENOSPC, &inode->i_mapping->flags))
10505                 ret = -ENOSPC;
10506         if (test_bit(AS_EIO, &inode->i_mapping->flags) &&
10507             test_and_clear_bit(AS_EIO, &inode->i_mapping->flags))
10508                 ret = -EIO;
10509
10510         return ret;
10511 }
10512
10513 static const struct inode_operations btrfs_dir_inode_operations = {
10514         .getattr        = btrfs_getattr,
10515         .lookup         = btrfs_lookup,
10516         .create         = btrfs_create,
10517         .unlink         = btrfs_unlink,
10518         .link           = btrfs_link,
10519         .mkdir          = btrfs_mkdir,
10520         .rmdir          = btrfs_rmdir,
10521         .rename2        = btrfs_rename2,
10522         .symlink        = btrfs_symlink,
10523         .setattr        = btrfs_setattr,
10524         .mknod          = btrfs_mknod,
10525         .setxattr       = generic_setxattr,
10526         .getxattr       = generic_getxattr,
10527         .listxattr      = btrfs_listxattr,
10528         .removexattr    = generic_removexattr,
10529         .permission     = btrfs_permission,
10530         .get_acl        = btrfs_get_acl,
10531         .set_acl        = btrfs_set_acl,
10532         .update_time    = btrfs_update_time,
10533         .tmpfile        = btrfs_tmpfile,
10534 };
10535 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10536         .lookup         = btrfs_lookup,
10537         .permission     = btrfs_permission,
10538         .get_acl        = btrfs_get_acl,
10539         .set_acl        = btrfs_set_acl,
10540         .update_time    = btrfs_update_time,
10541 };
10542
10543 static const struct file_operations btrfs_dir_file_operations = {
10544         .llseek         = generic_file_llseek,
10545         .read           = generic_read_dir,
10546         .iterate_shared = btrfs_real_readdir,
10547         .unlocked_ioctl = btrfs_ioctl,
10548 #ifdef CONFIG_COMPAT
10549         .compat_ioctl   = btrfs_compat_ioctl,
10550 #endif
10551         .release        = btrfs_release_file,
10552         .fsync          = btrfs_sync_file,
10553 };
10554
10555 static const struct extent_io_ops btrfs_extent_io_ops = {
10556         .fill_delalloc = run_delalloc_range,
10557         .submit_bio_hook = btrfs_submit_bio_hook,
10558         .merge_bio_hook = btrfs_merge_bio_hook,
10559         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10560         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10561         .writepage_start_hook = btrfs_writepage_start_hook,
10562         .set_bit_hook = btrfs_set_bit_hook,
10563         .clear_bit_hook = btrfs_clear_bit_hook,
10564         .merge_extent_hook = btrfs_merge_extent_hook,
10565         .split_extent_hook = btrfs_split_extent_hook,
10566 };
10567
10568 /*
10569  * btrfs doesn't support the bmap operation because swapfiles
10570  * use bmap to make a mapping of extents in the file.  They assume
10571  * these extents won't change over the life of the file and they
10572  * use the bmap result to do IO directly to the drive.
10573  *
10574  * the btrfs bmap call would return logical addresses that aren't
10575  * suitable for IO and they also will change frequently as COW
10576  * operations happen.  So, swapfile + btrfs == corruption.
10577  *
10578  * For now we're avoiding this by dropping bmap.
10579  */
10580 static const struct address_space_operations btrfs_aops = {
10581         .readpage       = btrfs_readpage,
10582         .writepage      = btrfs_writepage,
10583         .writepages     = btrfs_writepages,
10584         .readpages      = btrfs_readpages,
10585         .direct_IO      = btrfs_direct_IO,
10586         .invalidatepage = btrfs_invalidatepage,
10587         .releasepage    = btrfs_releasepage,
10588         .set_page_dirty = btrfs_set_page_dirty,
10589         .error_remove_page = generic_error_remove_page,
10590 };
10591
10592 static const struct address_space_operations btrfs_symlink_aops = {
10593         .readpage       = btrfs_readpage,
10594         .writepage      = btrfs_writepage,
10595         .invalidatepage = btrfs_invalidatepage,
10596         .releasepage    = btrfs_releasepage,
10597 };
10598
10599 static const struct inode_operations btrfs_file_inode_operations = {
10600         .getattr        = btrfs_getattr,
10601         .setattr        = btrfs_setattr,
10602         .setxattr       = generic_setxattr,
10603         .getxattr       = generic_getxattr,
10604         .listxattr      = btrfs_listxattr,
10605         .removexattr    = generic_removexattr,
10606         .permission     = btrfs_permission,
10607         .fiemap         = btrfs_fiemap,
10608         .get_acl        = btrfs_get_acl,
10609         .set_acl        = btrfs_set_acl,
10610         .update_time    = btrfs_update_time,
10611 };
10612 static const struct inode_operations btrfs_special_inode_operations = {
10613         .getattr        = btrfs_getattr,
10614         .setattr        = btrfs_setattr,
10615         .permission     = btrfs_permission,
10616         .setxattr       = generic_setxattr,
10617         .getxattr       = generic_getxattr,
10618         .listxattr      = btrfs_listxattr,
10619         .removexattr    = generic_removexattr,
10620         .get_acl        = btrfs_get_acl,
10621         .set_acl        = btrfs_set_acl,
10622         .update_time    = btrfs_update_time,
10623 };
10624 static const struct inode_operations btrfs_symlink_inode_operations = {
10625         .readlink       = generic_readlink,
10626         .get_link       = page_get_link,
10627         .getattr        = btrfs_getattr,
10628         .setattr        = btrfs_setattr,
10629         .permission     = btrfs_permission,
10630         .setxattr       = generic_setxattr,
10631         .getxattr       = generic_getxattr,
10632         .listxattr      = btrfs_listxattr,
10633         .removexattr    = generic_removexattr,
10634         .update_time    = btrfs_update_time,
10635 };
10636
10637 const struct dentry_operations btrfs_dentry_operations = {
10638         .d_delete       = btrfs_dentry_delete,
10639         .d_release      = btrfs_dentry_release,
10640 };